diff --git a/.gitignore b/.gitignore
index 9622ab78e0e0556ec2b4cc974fee93ff680d54d2..4f21fefda9f64a0392881971a715b97c234030e3 100644
--- a/.gitignore
+++ b/.gitignore
@@ -22,6 +22,7 @@ cmake-build-*
 
 # generated while compiling
 python/paddle/v2/framework/core.so
+paddle/pybind/pybind.h
 CMakeFiles
 cmake_install.cmake
 paddle/.timestamp
diff --git a/CMakeLists.txt b/CMakeLists.txt
index 08237cd850ae20c515a39c8783a18deaac431626..5739c2a26039426ab544f762e9401445f01e7de7 100644
--- a/CMakeLists.txt
+++ b/CMakeLists.txt
@@ -67,6 +67,9 @@ endif()
 if(ANDROID)
     if(${CMAKE_SYSTEM_VERSION} VERSION_LESS "16")
         message(FATAL_ERROR "Unsupport standalone toolchains with Android API level lower than 16")
+    elseif(${CMAKE_SYSTEM_VERSION} VERSION_LESS "21")
+        # TODO: support glog for Android api 16 ~ 19 in the future
+        message(WARNING "Using the unofficial git repository <https://github.com/Xreki/glog.git> instead")
     endif()
 
     set(WITH_GPU OFF CACHE STRING
diff --git a/Dockerfile.android b/Dockerfile.android
index 452aa1574550627c2cce6375e12e154a9763254d..9d13a414f67be04e17b7d83403228d92bce0eda9 100644
--- a/Dockerfile.android
+++ b/Dockerfile.android
@@ -6,13 +6,14 @@ RUN /bin/bash -c 'if [[ -n ${UBUNTU_MIRROR} ]]; then sed -i 's#http://archive.ub
 
 # ENV variables
 ARG ANDROID_ABI
+ARG ANDROID_API
 
 ENV ANDROID_ABI=${ANDROID_ABI:-"armeabi-v7a"}
+ENV ANDROID_API=${ANDROID_API:-21}
 
 ENV HOME=/root \
     ANDROID_NDK_HOME=/opt/android-ndk-linux \
-    ANDROID_ARM_STANDALONE_TOOLCHAIN=/opt/arm-toolchain \
-    ANDROID_ARM64_STANDALONE_TOOLCHAIN=/opt/arm64-toolchain
+    ANDROID_TOOLCHAINS_DIR=/opt/toolchains
 
 RUN apt-get update && \
     apt-get install -y \
@@ -42,14 +43,12 @@ RUN pip install --upgrade pip && \
     pip install pre-commit
 
 # Android NDK
-RUN mkdir /opt/android-ndk-tmp && \
+RUN mkdir -p ${ANDROID_TOOLCHAINS_DIR} && \
+    mkdir -p /opt/android-ndk-tmp && \
     cd /opt/android-ndk-tmp && \
     wget -q https://dl.google.com/android/repository/android-ndk-r14b-linux-x86_64.zip && \
     unzip -q android-ndk-r14b-linux-x86_64.zip && \
     mv android-ndk-r14b ${ANDROID_NDK_HOME} && \
-    ${ANDROID_NDK_HOME}/build/tools/make-standalone-toolchain.sh --arch=arm --platform=android-23 --install-dir=${ANDROID_ARM_STANDALONE_TOOLCHAIN} && \
-    ${ANDROID_NDK_HOME}/build/tools/make-standalone-toolchain.sh --arch=arm64 --platform=android-23 --install-dir=${ANDROID_ARM64_STANDALONE_TOOLCHAIN} && \
-    rm -rf /opt/android-ndk-tmp && \
-    rm -rf ${ANDROID_NDK_HOME}
+    rm -rf /opt/android-ndk-tmp
 
 CMD ["bash", "/paddle/paddle/scripts/docker/build_android.sh"]
diff --git a/cmake/cpplint.cmake b/cmake/cpplint.cmake
index 8d5d533126c9b7fa84c725d614cf3486126d0284..4823dc3e91390002aefac70f7931b4197db05789 100644
--- a/cmake/cpplint.cmake
+++ b/cmake/cpplint.cmake
@@ -26,9 +26,9 @@ set(IGNORE_PATTERN
     .*ImportanceSampler.*
     .*cblas\\.h.*
     .*\\.pb\\.txt
-    .*LtrDataProvider.*
     .*MultiDataProvider.*
-    .*pb.*)
+    .*pb.*
+    .*pybind.h)
 
 # add_style_check_target
 #
diff --git a/cmake/external/gflags.cmake b/cmake/external/gflags.cmake
index 16e5bef4cdb8d6513de51838e3c3c8398dbad60d..01a2f4d5fa357ca882162247cc52299a3d1d3030 100644
--- a/cmake/external/gflags.cmake
+++ b/cmake/external/gflags.cmake
@@ -18,9 +18,9 @@ SET(GFLAGS_SOURCES_DIR ${THIRD_PARTY_PATH}/gflags)
 SET(GFLAGS_INSTALL_DIR ${THIRD_PARTY_PATH}/install/gflags)
 SET(GFLAGS_INCLUDE_DIR "${GFLAGS_INSTALL_DIR}/include" CACHE PATH "gflags include directory." FORCE)
 IF(WIN32)
-    set(GFLAGS_LIBRARIES "${GFLAGS_INSTALL_DIR}/lib/gflags.lib" CACHE FILEPATH "GFLAGS_LIBRARIES" FORCE)
+  set(GFLAGS_LIBRARIES "${GFLAGS_INSTALL_DIR}/lib/gflags.lib" CACHE FILEPATH "GFLAGS_LIBRARIES" FORCE)
 ELSE(WIN32)
-    set(GFLAGS_LIBRARIES "${GFLAGS_INSTALL_DIR}/lib/libgflags.a" CACHE FILEPATH "GFLAGS_LIBRARIES" FORCE)
+  set(GFLAGS_LIBRARIES "${GFLAGS_INSTALL_DIR}/lib/libgflags.a" CACHE FILEPATH "GFLAGS_LIBRARIES" FORCE)
 ENDIF(WIN32)
 
 INCLUDE_DIRECTORIES(${GFLAGS_INCLUDE_DIR})
@@ -56,3 +56,12 @@ SET_PROPERTY(TARGET gflags PROPERTY IMPORTED_LOCATION ${GFLAGS_LIBRARIES})
 ADD_DEPENDENCIES(gflags extern_gflags)
 
 LIST(APPEND external_project_dependencies gflags)
+
+IF(WITH_C_API)
+  INSTALL(DIRECTORY ${GFLAGS_INCLUDE_DIR} DESTINATION third_party/gflags)
+  IF(ANDROID)
+    INSTALL(FILES ${GFLAGS_LIBRARIES} DESTINATION third_party/gflags/lib/${ANDROID_ABI})
+  ELSE()
+    INSTALL(FILES ${GFLAGS_LIBRARIES} DESTINATION third_party/gflags/lib)
+  ENDIF()
+ENDIF()
diff --git a/cmake/external/glog.cmake b/cmake/external/glog.cmake
index 8a594a825abdca6a0f989b94fa42f97d6df5e10a..b450a3016667dcb4ab229fe7ec8aaae8609d8171 100644
--- a/cmake/external/glog.cmake
+++ b/cmake/external/glog.cmake
@@ -19,9 +19,9 @@ SET(GLOG_INSTALL_DIR ${THIRD_PARTY_PATH}/install/glog)
 SET(GLOG_INCLUDE_DIR "${GLOG_INSTALL_DIR}/include" CACHE PATH "glog include directory." FORCE)
 
 IF(WIN32)
-    SET(GLOG_LIBRARIES "${GLOG_INSTALL_DIR}/lib/libglog.lib" CACHE FILEPATH "glog library." FORCE)
+  SET(GLOG_LIBRARIES "${GLOG_INSTALL_DIR}/lib/libglog.lib" CACHE FILEPATH "glog library." FORCE)
 ELSE(WIN32)
-    SET(GLOG_LIBRARIES "${GLOG_INSTALL_DIR}/lib/libglog.a" CACHE FILEPATH "glog library." FORCE)
+  SET(GLOG_LIBRARIES "${GLOG_INSTALL_DIR}/lib/libglog.a" CACHE FILEPATH "glog library." FORCE)
 ENDIF(WIN32)
 
 INCLUDE_DIRECTORIES(${GLOG_INCLUDE_DIR})
@@ -56,3 +56,12 @@ ADD_DEPENDENCIES(glog extern_glog gflags)
 LINK_LIBRARIES(glog gflags)
 
 LIST(APPEND external_project_dependencies glog)
+
+IF(WITH_C_API)
+  INSTALL(DIRECTORY ${GLOG_INCLUDE_DIR} DESTINATION third_party/glog)
+  IF(ANDROID)
+    INSTALL(FILES ${GLOG_LIBRARIES} DESTINATION third_party/glog/lib/${ANDROID_ABI})
+  ELSE()
+    INSTALL(FILES ${GLOG_LIBRARIES} DESTINATION third_party/glog/lib)
+  ENDIF()
+ENDIF()
diff --git a/cmake/external/openblas.cmake b/cmake/external/openblas.cmake
index f9e05af59fed7a8ad049390eda2c94d8577db1e7..4fc8d43fc10891603b79c01a1c769cae21c52655 100644
--- a/cmake/external/openblas.cmake
+++ b/cmake/external/openblas.cmake
@@ -73,6 +73,26 @@ IF(NOT ${CBLAS_FOUND})
         UPDATE_COMMAND      ""
         CONFIGURE_COMMAND   ""
     )
+
+    IF(WITH_C_API)
+        INSTALL(DIRECTORY ${CBLAS_INC_DIR} DESTINATION third_party/openblas)
+        # Because libopenblas.a is a symbolic link of another library, thus need to
+        # install the whole directory.
+        IF(ANDROID)
+            SET(TMP_INSTALL_DIR third_party/openblas/lib/${ANDROID_ABI})
+        ELSE()
+            SET(TMP_INSTALL_DIR third_party/openblas/lib)
+        ENDIF()
+        INSTALL(CODE "execute_process(
+            COMMAND ${CMAKE_COMMAND} -E copy_directory ${CBLAS_INSTALL_DIR}/lib
+                    destination ${CMAKE_INSTALL_PREFIX}/${TMP_INSTALL_DIR}
+            )"
+        )
+        INSTALL(CODE "MESSAGE(STATUS \"Installing: \"
+                \"${CBLAS_INSTALL_DIR}/lib -> ${CMAKE_INSTALL_PREFIX}/${TMP_INSTALL_DIR}\"
+            )"
+        )
+    ENDIF()
 ENDIF(NOT ${CBLAS_FOUND})
 
 MESSAGE(STATUS "BLAS library: ${CBLAS_LIBRARIES}")
diff --git a/cmake/external/protobuf.cmake b/cmake/external/protobuf.cmake
index e629d61585c2d2ff916187ee28d4fd089a5bd857..a887be2e2ae5e21562fc15c775bb24cc1553480e 100644
--- a/cmake/external/protobuf.cmake
+++ b/cmake/external/protobuf.cmake
@@ -223,6 +223,15 @@ IF(NOT PROTOBUF_FOUND)
     SET(PROTOBUF_PROTOC_LIBRARY ${extern_protobuf_PROTOC_LIBRARY}
         CACHE FILEPATH "protoc library." FORCE)
 
+    IF(WITH_C_API)
+        INSTALL(DIRECTORY ${PROTOBUF_INCLUDE_DIR} DESTINATION third_party/protobuf)
+        IF(ANDROID)
+            INSTALL(FILES ${PROTOBUF_LIBRARY} DESTINATION third_party/protobuf/lib/${ANDROID_ABI})
+        ELSE()
+            INSTALL(FILES ${PROTOBUF_LIBRARY} DESTINATION third_party/protobuf/lib)
+        ENDIF()
+    ENDIF()
+
     IF(CMAKE_CROSSCOMPILING)
         PROMPT_PROTOBUF_LIB(protobuf_host extern_protobuf)
     ELSE()
diff --git a/cmake/external/zlib.cmake b/cmake/external/zlib.cmake
index 45ca5542b7dc30216b45487782f849b93c5f8fca..5aecab90ca3cecdfdba0eac178a6ba07dfcb8745 100644
--- a/cmake/external/zlib.cmake
+++ b/cmake/external/zlib.cmake
@@ -49,3 +49,12 @@ ExternalProject_Add(
 )
 
 LIST(APPEND external_project_dependencies zlib)
+
+IF(WITH_C_API)
+  INSTALL(DIRECTORY ${ZLIB_INCLUDE_DIR} DESTINATION third_party/zlib)
+  IF(ANDROID)
+    INSTALL(FILES ${ZLIB_LIBRARIES} DESTINATION third_party/zlib/lib/${ANDROID_ABI})
+  ELSE()
+    INSTALL(FILES ${ZLIB_LIBRARIES} DESTINATION third_party/zlib/lib)
+  ENDIF()
+ENDIF()
diff --git a/doc/design/if_else_op.md b/doc/design/if_else_op.md
index 7370c2a24fa644a64e738f202bac9b9209642e08..954a19c0733358c235eae3cffe134c23dac94c95 100644
--- a/doc/design/if_else_op.md
+++ b/doc/design/if_else_op.md
@@ -1,22 +1,4 @@
-IfOp should have only one branch. An IfOp operator takes a `cond` variable whose value must be a vector of N boolean elements. Its return value has M (M<=N) instances, each corresponds to a true element in `cond`.
-
-```python
-import paddle as pd
-
-x = var()
-y = var()
-cond = var()
-
-b = pd.create_ifop(inputs=[x], output_num=1)
-with b.true_block():
-    x = b.inputs(0)
-    z = operator.add(x, y)
-    b.set_output(0, operator.softmax(z))
-
-out = b(cond)
-```
-
-If we want the output still has N instances, we can use IfElseOp with a default value, whose minibatch size must be N:
+IfOp should have only one branch. An IfOp operator takes a `cond` variable whose value must be a vector of N boolean elements. Its return value has N instances. If cond[i] == True, input instance input[i] will go through true_block() and generate output[i]; otherwise it will produce output from false_bloack().
 
 ```python
 import paddle as pd
@@ -39,7 +21,7 @@ with b.false_block():
 out = b(cond)
 ```
 
-If only true_block is set in an IfElseOp, we can have a default value for false as:
+If only true_block is set in an IfElseOp, a special case is that we can have a default value for false as:
 ```python
 import paddle as pd
 
diff --git a/doc/design/ops/images/2_level_rnn.dot b/doc/design/ops/images/2_level_rnn.dot
new file mode 100644
index 0000000000000000000000000000000000000000..a498e882a3d85a33d44dbad7474fa2a340e33976
--- /dev/null
+++ b/doc/design/ops/images/2_level_rnn.dot
@@ -0,0 +1,56 @@
+digraph G {
+
+  rnn [label="1-th level RNN" shape=box]
+
+  subgraph cluster0 {
+    label = "time step 0"
+
+    sent0 [label="sentence"]
+    sent1 [label="sentence"]
+
+    rnn1 [label="2-th level RNN" shape=box]
+
+    sent0 -> rnn1
+    sent1 -> rnn1
+  }
+
+  subgraph cluster1 {
+    label = "time step 1"
+
+    sent2 [label="sentence"]
+    sent3 [label="sentence"]
+
+    rnn2 [label="2-th level RNN" shape=box]
+
+    sent2 -> rnn2
+    sent3 -> rnn2
+  }
+
+  subgraph cluster2 {
+    label = "time step 2"
+
+    sent4 [label="sentence"]
+    sent5 [label="sentence"]
+
+    rnn3 [label="2-th level RNN" shape=box]
+
+    sent4 -> rnn3
+    sent5 -> rnn3
+  }
+
+
+  para0 [label="paragraph info 0"]
+  para1 [label="paragraph info 1"]
+  para2 [label="paragraph info 2"]
+
+  rnn1 -> para0
+  rnn2 -> para1
+  rnn3 -> para2
+
+  para0 -> rnn
+  para1 -> rnn
+  para2 -> rnn
+
+  chapter [label="chapter info"]
+  rnn -> chapter
+}
diff --git a/doc/design/ops/images/2_level_rnn.png b/doc/design/ops/images/2_level_rnn.png
new file mode 100644
index 0000000000000000000000000000000000000000..0537a75beb175c0c284717421f7aa908da2a5038
Binary files /dev/null and b/doc/design/ops/images/2_level_rnn.png differ
diff --git a/doc/design/ops/images/rnn.dot b/doc/design/ops/images/rnn.dot
new file mode 100644
index 0000000000000000000000000000000000000000..c1141cd9c981bb3cbf50d8bf7a6ed210280d79a5
--- /dev/null
+++ b/doc/design/ops/images/rnn.dot
@@ -0,0 +1,87 @@
+digraph G {
+  label = "simple RNN implementation" 
+
+  ranksep=2;
+
+  //graph [nodesep=1, ranksep=1];
+
+  node[nodesep=1]
+
+  subgraph cluster0 {
+    label = "global scope"
+    rankdir = TB
+    W
+    boot_memory
+    input
+    output
+  }
+
+  subgraph cluster1 {
+    label = "step-scope 0"
+    rankdir = TB
+    memory0[label="memory"]
+    prememory0[label="pre-memory"]
+    step_input0[label="step input"]
+    step_output0[label="step output"]
+  }
+
+  subgraph cluster2 {
+    label = "step-scope 1"
+    rankdir = TB
+    memory1[label="memory"]
+    prememory1[label="pre-memory"]
+    step_input1[label="step input"]
+    step_output1[label="step output"]
+  }
+
+  subgraph cluster3 {
+    label = "step-scope 2"
+    rankdir = TB
+    memory2[label="memory"]
+    prememory2[label="pre-memory"]
+    step_input2[label="step input"]
+    step_output2[label="step output"]
+  }
+
+  stepnet [shape=box]
+  stepnet0 [shape=box, style=dashed]
+  stepnet1 [shape=box, style=dashed]
+  stepnet2 [shape=box, style=dashed]
+
+
+  edge[color=blue]
+  boot_memory -> prememory0 [label="init" color="blue"]
+  memory0 -> prememory1  [label="copy/reference" color="blue"]
+  memory1 -> prememory2 [label="copy/reference" color="blue"]
+
+  edge[color=black]
+  W -> stepnet0[constraint=false, style=dashed]
+  W -> stepnet1[constraint=false, style=dashed]
+  W -> stepnet2[constraint=false, style=dashed]
+
+  memory0 -> stepnet0[style=dashed]
+  prememory0 -> stepnet0 -> step_output0[style=dashed]
+
+  memory1 -> stepnet1[style=dashed]
+  prememory1 -> stepnet1 -> step_output1[style=dashed]
+
+  memory2 -> stepnet2[style=dashed]
+  prememory2 -> stepnet2 -> step_output2[style=dashed]
+
+  input -> step_input0
+  input -> step_input1
+  input -> step_input2
+
+  step_input0 -> stepnet0 [style=dashed]
+  step_input1 -> stepnet1[style=dashed]
+  step_input2 -> stepnet2[style=dashed]
+
+  step_output0 -> output
+  step_output1 -> output
+  step_output2 -> output
+
+  stepnet0 -> stepnet[style=dashed]
+  stepnet1 -> stepnet[style=dashed]
+  stepnet2 -> stepnet[style=dashed]
+
+}
diff --git a/doc/design/ops/images/rnn.jpg b/doc/design/ops/images/rnn.jpg
new file mode 100644
index 0000000000000000000000000000000000000000..9867e404cf959df0dce6ded5222b466c788fb840
Binary files /dev/null and b/doc/design/ops/images/rnn.jpg differ
diff --git a/doc/design/ops/images/rnn.png b/doc/design/ops/images/rnn.png
new file mode 100644
index 0000000000000000000000000000000000000000..e139e373fe8396782044cfd936fdde624f8c66fe
Binary files /dev/null and b/doc/design/ops/images/rnn.png differ
diff --git a/doc/design/ops/images/rnn_2level_data.dot b/doc/design/ops/images/rnn_2level_data.dot
new file mode 100644
index 0000000000000000000000000000000000000000..1d85ae2617a915ad0ad8288d848b607cc37ad297
--- /dev/null
+++ b/doc/design/ops/images/rnn_2level_data.dot
@@ -0,0 +1,75 @@
+digraph G {
+  chapter [label="chapter"]
+
+  subgraph cluster0 {
+    label = "paragraph 0"
+
+    top_rnn0[label="top rnn step 0" shape=box]
+
+    p0 [label="paragraph 0"]
+    p1 [label="paragraph 1"]
+  }
+
+  subgraph cluster1{
+    label = "paragraph 1"
+
+    top_rnn1[label="top rnn step 1" shape=box]
+
+    p2 [label="paragraph 0"]
+    p3 [label="paragraph 1"]
+  }
+
+  subgraph cluster_p0 {
+    label = "sentence 0"
+
+    low_rnn0 [label="low rnn step 0" shape=box]
+    s00 [label="sentence 0"]
+    s01 [label="sentence 1"]
+
+    low_rnn0 -> s00
+    low_rnn0 -> s01
+  }
+
+  subgraph cluster_p1 {
+    label = "sentence 1"
+    low_rnn1 [label="low rnn step 1" shape=box]
+    s10 [label="sentence 0"]
+    s11 [label="sentence 1"]
+    low_rnn1 -> s10
+    low_rnn1 -> s11
+  }
+
+  subgraph cluster_p2 {
+    label = "sentence 1"
+    low_rnn2 [label="low rnn step 0" shape=box]
+    s20 [label="sentence 0"]
+    s21 [label="sentence 1"]
+    low_rnn2 -> s20
+    low_rnn2 -> s21
+  }
+
+  subgraph cluster_p3 {
+    label = "sentence 1"
+    low_rnn3 [label="low rnn step 1" shape=box]
+    s30 [label="sentence 0"]
+    s31 [label="sentence 1"]
+    low_rnn3 -> s30
+    low_rnn3 -> s31
+  }
+
+
+  chapter -> top_rnn0
+  chapter -> top_rnn1
+
+  top_rnn0 -> p0
+  top_rnn0 -> p1
+  top_rnn1 -> p2
+  top_rnn1 -> p3
+
+
+  p0 -> low_rnn0
+  p1 -> low_rnn1
+  p2 -> low_rnn2
+  p3 -> low_rnn3
+
+}
diff --git a/doc/design/ops/images/rnn_2level_data.png b/doc/design/ops/images/rnn_2level_data.png
new file mode 100644
index 0000000000000000000000000000000000000000..4be81b2430717a6a506342a09fc26899568574c6
Binary files /dev/null and b/doc/design/ops/images/rnn_2level_data.png differ
diff --git a/doc/design/ops/rnn.md b/doc/design/ops/rnn.md
new file mode 100644
index 0000000000000000000000000000000000000000..a78eea7d45e9e9553d153170aa31da55ec6e8289
--- /dev/null
+++ b/doc/design/ops/rnn.md
@@ -0,0 +1,153 @@
+# RNNOp design
+
+This document is about an RNN operator which requires that instances in a mini-batch have the same length.  We will have a more flexible RNN operator.
+
+## RNN Algorithm Implementation
+
+<p aligh="center">
+<img src="./images/rnn.jpg"/>
+</p>
+
+The above diagram shows an RNN unrolled into a full network.
+
+There are several important concepts:
+
+- *step-net*: the sub-graph to run at each step,
+- *memory*, $h_t$, the state of the current step,
+- *ex-memory*, $h_{t-1}$, the state of the previous step,
+- *initial memory value*, the ex-memory of the first step.
+
+### Step-scope
+
+There could be local variables defined in step-nets.  PaddlePaddle runtime realizes these variables in *step-scopes* -- scopes created for each step.
+
+<p aligh="center">
+<img src="./images/rnn.png"/><br/>
+Figure 2 the RNN's data flow
+</p>
+
+Please be aware that all steps run the same step-net.  Each step
+
+1. creates the step-scope,
+2. realizes local variables, including step-outputs, in the step-scope, and
+3. runs the step-net, which could use these variables.
+
+The RNN operator will compose its output from step outputs in step scopes.
+
+### Memory and Ex-memory
+
+Let's give more details about memory and ex-memory via a simply example:
+
+$$
+h_t = U h_{t-1} + W x_t
+$$,
+
+where $h_t$ and $h_{t-1}$ are the memory and ex-memory of step $t$'s respectively.
+
+In the implementation, we can make an ex-memory variable either "refers to" the memory variable of the previous step,
+or copy the value of the previous memory value to the current ex-memory variable.
+
+### Usage in Python
+
+For more information on Block, please refer to the [design doc](https://github.com/PaddlePaddle/Paddle/blob/develop/doc/design/block.md).
+
+We can define an RNN's step-net using Block:
+
+```python
+import paddle as pd
+
+X = some_op() # x is some operator's output, and is a LoDTensor
+a = some_op()
+
+# declare parameters
+W = pd.Variable(shape=[20, 30])
+U = pd.Variable(shape=[20, 30])
+
+rnn = pd.create_rnn_op(output_num=1)
+with rnn.stepnet():
+    x = rnn.add_input(X)
+    # declare a memory (rnn's step)
+    h = rnn.add_memory(init=a)
+    # h.pre_state() means previous memory of rnn
+    new_state = pd.add_two( pd.matmul(W, x) + pd.matmul(U, h.pre_state()))
+    # update current memory
+    h.update(new_state)
+    # indicate that h variables in all step scopes should be merged
+    rnn.add_outputs(h)
+
+out = rnn()
+```
+
+Python API functions in above example:
+
+- `rnn.add_input` indicates the parameter is a variable that will be segmented into step-inputs.
+- `rnn.add_memory` creates a variable used as the memory.
+- `rnn.add_outputs` mark the variables that will be concatenated across steps into the RNN output.
+
+### Nested RNN and LoDTensor
+
+An RNN whose step-net includes other RNN operators is known as an *nested RNN*.
+
+For example, we could have a 2-level RNN, where the top level corresponds to paragraphs, and the lower level corresponds to sentences.
+
+The following figure illustrates the feeding of text into the lower level, one sentence each step, and the feeding of step outputs to the top level. The final top level output is about the whole text.
+
+<p aligh="center">
+<img src="./images/2_level_rnn.png"/>
+</p>
+
+```python
+import paddle as pd
+
+W = pd.Variable(shape=[20, 30])
+U = pd.Variable(shape=[20, 30])
+
+W0 = pd.Variable(shape=[20, 30])
+U0 = pd.Variable(shape=[20, 30])
+
+# a is output of some op
+a = some_op()
+
+# chapter_data is a set of 128-dim word vectors
+# the first level of LoD is sentence
+# the second level of LoD is chapter
+chapter_data = pd.Variable(shape=[None, 128], type=pd.lod_tensor, level=2)
+
+def lower_level_rnn(paragraph):
+    '''
+    x: the input
+    '''
+    rnn = pd.create_rnn_op(output_num=1)
+    with rnn.stepnet():
+        sentence = rnn.add_input(paragraph, level=0)
+        h = rnn.add_memory(shape=[20, 30])
+        h.update(
+            pd.matmul(W, sentence) + pd.matmul(U, h.pre_state()))
+        # get the last state as sentence's info
+        rnn.add_outputs(h)
+    return rnn
+
+top_level_rnn = pd.create_rnn_op(output_num=1)
+with top_level_rnn.stepnet():
+    paragraph_data = rnn.add_input(chapter_data, level=1)
+    low_rnn = lower_level_rnn(paragraph_data)
+    paragraph_out = low_rnn()
+
+    h = rnn.add_memory(init=a)
+    h.update(
+        pd.matmul(W0, paragraph_data) + pd.matmul(U0, h.pre_state()))
+    top_level_rnn.add_outputs(h)
+
+# just output the last step
+chapter_out = top_level_rnn(output_all_steps=False)
+```
+
+in above example, the construction of the `top_level_rnn` calls  `lower_level_rnn`.  The input is a LoD Tensor. The top level RNN segments input text data into paragraphs, and the lower level RNN segments each paragraph into sentences.
+
+By default, the `RNNOp` will concatenate the outputs from all the time steps,
+if the `output_all_steps` set to False, it will only output the final time step.
+
+
+<p align="center">
+<img src="images/rnn_2level_data.png"/>
+</p>
diff --git a/doc/howto/dev/new_op_cn.md b/doc/howto/dev/new_op_cn.md
index e3892849abe21fc207d2fcbe4adc65184ba771f4..c6570b89aedfaac1aef9b00e889b0b3ed21d8d65 100644
--- a/doc/howto/dev/new_op_cn.md
+++ b/doc/howto/dev/new_op_cn.md
@@ -34,7 +34,7 @@ Kernel实现       | CPU、GPU共享Kernel实现在`.h`文件中，否则，CPU
 注册Op           | Op注册实现在`.cc`文件；Kernel注册CPU实现在`.cc`文件中，GPU实现在`.cu`文件中
 
 
-实现新的op都添加至目录[paddle/operators](https://github.com/PaddlePaddle/Paddle/tree/develop/paddle/operators)下，文件命名以`*_op.h`（如有） 、 `*_op.cc` 、`*_op.cu`（如有）结尾。
+实现新的op都添加至目录[paddle/operators](https://github.com/PaddlePaddle/Paddle/tree/develop/paddle/operators)下，文件命名以`*_op.h`（如有） 、 `*_op.cc` 、`*_op.cu`（如有）结尾。**系统会根据文件名自动构建op和其对应的Python扩展。**
 
 
 下面以矩阵乘操作，即[MulOp](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/operators/mul_op.cc)为例来介绍如何写带Kernel的Operator。
@@ -224,45 +224,15 @@ MulOp(const std::string &type, const framework::VariableNameMap &inputs,
 
 ### 5. 编译
 
-- 简单**无特殊依赖**的OP无需修改CMakeList.txt文件。[paddle/operators/CMakeLists.txt](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/operators/CMakeLists.txt) 会自动将 `paddle/operators` 目录下新增的 `*_op.cc` 文件加入编译。
-- 较为复杂、**有额外依赖** 的operator仍需要修改[paddle/operators/CMakeLists.txt](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/operators/CMakeLists.txt)。如，`mul_op` 依赖 `math_function`，需要在`CMakeLists.txt`中添加如下内容：
+运行下面命令可以进行编译：
 
-    ```
-    op_library(mul_op SRCS mul_op.cc mul_op.cu DEPS math_function)		 +
-    ```
-
-- 运行下面命令可以进行编译：
-
-    ```
-    make mul_op
-    ```
+```
+make mul_op
+```
 
 ## 绑定Python
 
-- 绑定Python
-
-    在 [`paddle/pybind/pybind.cc
-`](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/pybind/pybind.cc) 使用`USE_OP`告知编译器需要链接的Op，具体解释参考[代码注释](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/framework/op_registry.h#L81)。
-
-    ```
-    USE_OP(mul);
-    ```
-    如果只实现了CPU版本，则使用`USE_CPU_ONLY_OP`:
-
-    ```
-    USE_CPU_ONLY_OP(gather);
-    ```
-
-    如果OP不带Kernel，则使用`USE_NO_KENREL_OP`:
-
-    ```
-    USE_NO_KENREL_OP(recurrent);
-    ```
-
-
- - 生成库
-
-   `paddle/operators` 目录下新增的 `*_op.cc` 文件会被自动添加链接到生成的lib库中。
+系统会对新增的op自动绑定Python，并链接到生成的lib库中。
 
 ## 实现单元测试
 
@@ -367,3 +337,10 @@ make test ARGS="-R test_mul_op -V"
 ```bash
 ctest -R test_mul_op
 ```
+
+## 注意事项
+
+- 为每个Op创建单独的`*_op.h`（如有）、`*_op.cc`和`*_op.cu`（如有）。不允许一个文件中包含多个Op，这将会导致编译出错。
+- 注册Op时的类型名，需要和该Op的名字一样。即不允许在`A_op.cc`里面，注册`REGISTER_OP(B, ...)`等，这将会导致单元测试出错。
+- 如果Op没有实现GPU Kernel，请不要创建空的`*_op.cu`，这将会导致单元测试出错。
+- 如果多个Op依赖一些共用的函数，可以创建非`*_op.*`格式的文件来存放，如`gather.h`文件。
diff --git a/paddle/capi/CMakeLists.txt b/paddle/capi/CMakeLists.txt
index dde99ab3400be4e61bfe119fc272270518acf070..3af111eb5738c3f2f399ff4e5c06c8d2ecd8973e 100644
--- a/paddle/capi/CMakeLists.txt
+++ b/paddle/capi/CMakeLists.txt
@@ -64,9 +64,29 @@ link_paddle_exe(paddle_capi_shared)
 install(FILES ${CAPI_HEADERS} DESTINATION include/paddle)
 install(FILES ${CMAKE_CURRENT_BINARY_DIR}/config.h DESTINATION include/paddle)
 if(ANDROID)
+  execute_process(
+    COMMAND ${GIT_EXECUTABLE} log --pretty=oneline -1
+    OUTPUT_VARIABLE GIT_COMMITS_LIST
+    RESULT_VARIABLE GIT_COMMITS_LIST_RESULT
+    ERROR_QUIET OUTPUT_STRIP_TRAILING_WHITESPACE)
+  if(${GIT_COMMITS_LIST_RESULT})
+    set(GIT_COMMITS_LIST "No commits.")
+  endif()
   install(FILES ${CMAKE_CURRENT_BINARY_DIR}/${capi_whole_library}
           DESTINATION lib/${ANDROID_ABI})
   install(TARGETS paddle_capi_shared DESTINATION lib/${ANDROID_ABI})
+  install(CODE "FILE(WRITE ${CMAKE_INSTALL_PREFIX}/lib/${ANDROID_ABI}/BUILD.txt
+          \"Compiler:\n\"
+          \"\\t${CMAKE_C_COMPILER}\\n\"
+          \"\\t${CMAKE_CXX_COMPILER}\\n\"
+          \"Compiler Flags:\\n\"
+          \"\\t${CMAKE_F_FLAGS}\\n\"
+          \"\\t${CMAKE_CXX_FLAGS}\\n\"
+          \"Android API: ${CMAKE_SYSTEM_VERSION}\\n\"
+          \"Lastest commit:\\n\"
+          \"\\t${GIT_COMMITS_LIST}\\n\"
+      )"
+  )
 else(ANDROID)
   install(FILES ${CMAKE_CURRENT_BINARY_DIR}/${capi_whole_library} DESTINATION lib)
   install(TARGETS paddle_capi_shared DESTINATION lib)
diff --git a/paddle/framework/lod_tensor.h b/paddle/framework/lod_tensor.h
index 568f4e89819c8345d8908634f6fa56f09483a763..fac5cd20aa7f9db0792f8102bb442192ab1ad63f 100644
--- a/paddle/framework/lod_tensor.h
+++ b/paddle/framework/lod_tensor.h
@@ -51,18 +51,15 @@ bool operator==(const LoD& a, const LoD& b);
  * LoDTensor (Level of details Tensor)
  * see https://en.wikipedia.org/wiki/Level_of_details for reference.
  */
-class LoDTensor {
+class LoDTensor : public Tensor {
  public:
   LoDTensor() {}
-  LoDTensor(const LoD& lod, Tensor* t) : lod_(lod), tensor_(t) {}
 
-  void set_lod(const LoD& lod) { lod_ = lod; }
-
-  void set_tensor(Tensor* tensor) { tensor_ = tensor; }
+  explicit LoDTensor(const LoD& lod) : lod_(lod) {}
 
-  Tensor& tensor() { return *tensor_; }
+  void set_lod(const LoD& lod) { lod_ = lod; }
 
-  LoD lod() { return lod_; }
+  LoD lod() const { return lod_; }
 
   /*
    * Get a element from LoD.
@@ -104,7 +101,6 @@ class LoDTensor {
 
  private:
   LoD lod_;
-  Tensor* tensor_;  // not owned
 };
 }  // namespace framework
 }  // namespace paddle
diff --git a/paddle/framework/lod_tensor_test.cc b/paddle/framework/lod_tensor_test.cc
index 1da8553134f377f7a4fbe8008d12fe8d4a0e47f4..7915326b27a22e9280e3f09d9bbfc2a58f46aff7 100644
--- a/paddle/framework/lod_tensor_test.cc
+++ b/paddle/framework/lod_tensor_test.cc
@@ -36,69 +36,64 @@ class LoDTensorTester : public ::testing::Test {
 
     ASSERT_EQ(lod.size(), 3UL);
 
-    tensor.Resize({20 /*batch size*/, 128 /*dim*/});
+    lod_tensor_.Resize({20 /*batch size*/, 128 /*dim*/});
     // malloc memory
-    tensor.mutable_data<float>(place);
+    lod_tensor_.mutable_data<float>(place);
 
-    lod_tensor.set_lod(lod);
-    lod_tensor.set_tensor(&tensor);
+    lod_tensor_.set_lod(lod);
   }
 
  protected:
   platform::CPUPlace place;
-  Tensor tensor;
-  LoDTensor lod_tensor;
+  LoDTensor lod_tensor_;
 };
 
-TEST_F(LoDTensorTester, NumLevels) { ASSERT_EQ(lod_tensor.NumLevels(), 3UL); }
+TEST_F(LoDTensorTester, NumLevels) { ASSERT_EQ(lod_tensor_.NumLevels(), 3UL); }
 
 TEST_F(LoDTensorTester, NumElements) {
-  ASSERT_EQ(lod_tensor.NumElements(0), 2UL);
-  ASSERT_EQ(lod_tensor.NumElements(1), 4UL);
-  ASSERT_EQ(lod_tensor.NumElements(2), 8UL);
+  ASSERT_EQ(lod_tensor_.NumElements(0), 2UL);
+  ASSERT_EQ(lod_tensor_.NumElements(1), 4UL);
+  ASSERT_EQ(lod_tensor_.NumElements(2), 8UL);
 }
 
 TEST_F(LoDTensorTester, SliceLevels) {
   // slice 1 level
   for (size_t level = 0; level < 3UL; ++level) {
-    LoDTensor new_lod_tensor = lod_tensor;
+    LoDTensor new_lod_tensor = lod_tensor_;
     new_lod_tensor.SliceLevels(level, level + 1);
     ASSERT_EQ(new_lod_tensor.NumLevels(), 1UL);
-    ASSERT_EQ(new_lod_tensor.NumElements(0), lod_tensor.NumElements(level));
-    ASSERT_EQ(new_lod_tensor.tensor().data<float>(),
-              lod_tensor.tensor().data<float>());
+    ASSERT_EQ(new_lod_tensor.NumElements(0), lod_tensor_.NumElements(level));
+    ASSERT_EQ(new_lod_tensor.data<float>(), lod_tensor_.data<float>());
   }
   // slice 2 level
   for (size_t level = 0; level < 2UL; ++level) {
-    LoDTensor new_lod_tensor = lod_tensor;
+    LoDTensor new_lod_tensor = lod_tensor_;
     new_lod_tensor.SliceLevels(level, level + 2);
     ASSERT_EQ(new_lod_tensor.NumLevels(), 2UL);
-    ASSERT_EQ(new_lod_tensor.NumElements(0), lod_tensor.NumElements(level));
-    ASSERT_EQ(new_lod_tensor.NumElements(1), lod_tensor.NumElements(level + 1));
-    ASSERT_EQ(new_lod_tensor.tensor().data<float>(),
-              lod_tensor.tensor().data<float>());
+    ASSERT_EQ(new_lod_tensor.NumElements(0), lod_tensor_.NumElements(level));
+    ASSERT_EQ(new_lod_tensor.NumElements(1),
+              lod_tensor_.NumElements(level + 1));
+    ASSERT_EQ(new_lod_tensor.data<float>(), lod_tensor_.data<float>());
   }
 }
 
 TEST_F(LoDTensorTester, SliceInLevel) {
   size_t level = 0;
-  LoDTensor new_lod_tensor = lod_tensor;
+  LoDTensor new_lod_tensor = lod_tensor_;
   new_lod_tensor.SliceInLevel(level, 0, 2);
   EXPECT_EQ(new_lod_tensor.NumLevels(), 3UL);
   EXPECT_EQ(new_lod_tensor.NumElements(0), 2UL);
   EXPECT_EQ(new_lod_tensor.NumElements(1), 4UL);
   EXPECT_EQ(new_lod_tensor.NumElements(2), 8UL);
-  ASSERT_EQ(new_lod_tensor.tensor().data<float>(),
-            lod_tensor.tensor().data<float>());
+  ASSERT_EQ(new_lod_tensor.data<float>(), lod_tensor_.data<float>());
 
   level = 1;
-  new_lod_tensor = lod_tensor;
+  new_lod_tensor = lod_tensor_;
   new_lod_tensor.SliceInLevel(level, 0, 2);
   ASSERT_EQ(new_lod_tensor.NumLevels(), 2UL);
   ASSERT_EQ(new_lod_tensor.NumElements(0), 2UL);
   ASSERT_EQ(new_lod_tensor.NumElements(1), 4UL);
-  ASSERT_EQ(new_lod_tensor.tensor().data<float>(),
-            lod_tensor.tensor().data<float>());
+  ASSERT_EQ(new_lod_tensor.data<float>(), lod_tensor_.data<float>());
 }
 
 }  // namespace framework
diff --git a/paddle/framework/lod_tensor_test.cu b/paddle/framework/lod_tensor_test.cu
index 1079a36a2e7b24f6f8a5bcbb296355567305a765..97e69cdb2e5e1e64031c899f5e04020665485ba8 100644
--- a/paddle/framework/lod_tensor_test.cu
+++ b/paddle/framework/lod_tensor_test.cu
@@ -26,18 +26,16 @@ __global__ void test(size_t* a, int size) {
 }
 
 TEST(LoDTensor, LoDInGPU) {
-  paddle::framework::Tensor tensor;
   paddle::framework::LoDTensor lod_tensor;
   paddle::platform::GPUPlace place(0);
 
   paddle::framework::LoD src_lod;
   src_lod.push_back(std::vector<size_t>{0, 2, 4, 6, 8, 10, 12, 14});
 
-  tensor.Resize({14, 16});
-  tensor.mutable_data<float>(place);
+  lod_tensor.Resize({14, 16});
+  lod_tensor.mutable_data<float>(place);
 
   lod_tensor.set_lod(src_lod);
-  lod_tensor.set_tensor(&tensor);
   CHECK_EQ(lod_tensor.lod_element(0, 2), 4);
   CHECK_EQ(lod_tensor.lod_element(0, 4), 8);
 
diff --git a/paddle/framework/operator.cc b/paddle/framework/operator.cc
index e1e122091f7759b1a68f1f982bc2a35e8241f9f0..c57537be4bf67a8db6a49669ab8d2ed1b1324bdc 100644
--- a/paddle/framework/operator.cc
+++ b/paddle/framework/operator.cc
@@ -186,6 +186,48 @@ void OperatorBase::GenerateTemporaryNames() {
   }
 }
 
+template <>
+const Tensor* InferShapeContext::Input<Tensor>(const std::string& name) const {
+  auto* var = InputVar(name);
+  return var == nullptr ? nullptr : GetTensorFromVar(var);
+}
+
+template <>
+const std::vector<const Tensor*> InferShapeContext::MultiInput<Tensor>(
+    const std::string& name) const {
+  auto names = op().Inputs(name);
+  std::vector<const Tensor*> res;
+  res.reserve(names.size());
+  std::transform(names.begin(), names.end(), std::back_inserter(res),
+                 [&](const std::string& sub_name) {
+                   auto var = scope_.FindVar(sub_name);
+                   return var == nullptr ? nullptr : GetTensorFromVar(var);
+                 });
+  return res;
+}
+
+template <>
+Tensor* ExecutionContext::Output<Tensor>(const std::string& name) const {
+  auto* var = OutputVar(name);
+  return var == nullptr ? nullptr : const_cast<Tensor*>(GetTensorFromVar(var));
+}
+
+template <>
+std::vector<Tensor*> ExecutionContext::MultiOutput<Tensor>(
+    const std::string& name) const {
+  auto names = op().Outputs(name);
+  std::vector<Tensor*> res;
+  res.reserve(names.size());
+  std::transform(names.begin(), names.end(), std::back_inserter(res),
+                 [&](const std::string& sub_name) {
+                   auto var = scope().FindVar(sub_name);
+                   return var == nullptr
+                              ? nullptr
+                              : const_cast<Tensor*>(GetTensorFromVar(var));
+                 });
+  return res;
+}
+
 void OpProtoAndCheckerMaker::Validate() {
   validated_ = true;
   CheckNoDuplicatedInOutAttrs();
diff --git a/paddle/framework/operator.h b/paddle/framework/operator.h
index 4600b06009bcef7d0774d25b816aac4733f30795..adae7bfc3d7d31b1ed0373f01db4ef80343a08f7 100644
--- a/paddle/framework/operator.h
+++ b/paddle/framework/operator.h
@@ -22,6 +22,7 @@ limitations under the License. */
 #include "op_info.h"
 #include "paddle/framework/attribute.h"
 #include "paddle/framework/framework.pb.h"
+#include "paddle/framework/lod_tensor.h"
 #include "paddle/framework/scope.h"
 #include "paddle/framework/tensor.h"
 #include "paddle/platform/device_context.h"
@@ -326,11 +327,27 @@ class InferShapeContext {
     return res;
   }
 
+  const Tensor* GetTensorFromVar(const Variable* var) const {
+    if (var->IsType<LoDTensor>()) {
+      return &var->Get<LoDTensor>();
+    }
+    PADDLE_ENFORCE(var->IsType<Tensor>(),
+                   "The Input(%s) must be LoDTensor or Tensor.");
+    return &var->Get<Tensor>();
+  }
+
  private:
   const OperatorBase& op_;
   const Scope& scope_;
 };
 
+template <>
+const Tensor* InferShapeContext::Input<Tensor>(const std::string& name) const;
+
+template <>
+const std::vector<const Tensor*> InferShapeContext::MultiInput<Tensor>(
+    const std::string& name) const;
+
 template <typename T>
 struct EigenDeviceConverter;
 
@@ -363,9 +380,37 @@ class ExecutionContext : public InferShapeContext {
     return device_context_;
   }
 
+  // redefine Output function,
+  // use Variable::Get instead of Variable::GetMutable
+  template <typename T>
+  T* Output(const std::string& name) const {
+    auto var = OutputVar(name);
+    return var == nullptr ? nullptr : const_cast<T*>(&var->Get<T>());
+  }
+
+  // redefine MultiOutput function.
+  // use Variable::Get instead of Variable::GetMutable
+  template <typename T>
+  std::vector<T*> MultiOutput(const std::string& name) const {
+    auto names = op().Outputs(name);
+    std::vector<T*> res;
+    res.reserve(names.size());
+    std::transform(
+        names.begin(), names.end(), std::back_inserter(res),
+        [&](const std::string& sub_name) { return Output<T>(sub_name); });
+    return res;
+  }
+
   const platform::DeviceContext* device_context_;
 };
 
+template <>
+Tensor* ExecutionContext::Output<Tensor>(const std::string& name) const;
+
+template <>
+std::vector<Tensor*> ExecutionContext::MultiOutput<Tensor>(
+    const std::string& name) const;
+
 class OpKernel {
  public:
   /**
diff --git a/paddle/framework/tensor_impl.h b/paddle/framework/tensor_impl.h
index 642b53efc7095d25712ca324638f5fe9b8316c0c..ed166935f76be9d25062b5e69536c7b7ac19045d 100644
--- a/paddle/framework/tensor_impl.h
+++ b/paddle/framework/tensor_impl.h
@@ -22,7 +22,7 @@ namespace framework {
 template <typename T>
 inline void Tensor::check_memory_size() const {
   PADDLE_ENFORCE_NOT_NULL(
-      holder_, "Tenosr holds no memory. Call Tensor::mutable_data first.");
+      holder_, "Tensor holds no memory. Call Tensor::mutable_data first.");
   PADDLE_ENFORCE_GE(
       holder_->size(), numel() * sizeof(T) + offset_,
       "Tensor's dims_ is out of bound. Call Tensor::mutable_data "
diff --git a/paddle/framework/tensor_test.cc b/paddle/framework/tensor_test.cc
index 55302ea47120f420e952b26830c8ea4cbcce6435..e2ec738de35c90c6a06c9a46b062d4cce55f5eda 100644
--- a/paddle/framework/tensor_test.cc
+++ b/paddle/framework/tensor_test.cc
@@ -36,7 +36,7 @@ TEST(Tensor, DataAssert) {
   } catch (paddle::platform::EnforceNotMet err) {
     caught = true;
     std::string msg =
-        "holder_ should not be null\nTenosr holds no memory. Call "
+        "holder_ should not be null\nTensor holds no memory. Call "
         "Tensor::mutable_data first.";
     const char* what = err.what();
     for (size_t i = 0; i < msg.length(); ++i) {
@@ -112,7 +112,7 @@ TEST(Tensor, ShareDataWith) {
     } catch (paddle::platform::EnforceNotMet err) {
       caught = true;
       std::string msg =
-          "holder_ should not be null\nTenosr holds no memory. Call "
+          "holder_ should not be null\nTensor holds no memory. Call "
           "Tensor::mutable_data first.";
       const char* what = err.what();
       for (size_t i = 0; i < msg.length(); ++i) {
@@ -274,4 +274,4 @@ TEST(Tensor, ReshapeToMatrix) {
   Tensor res = ReshapeToMatrix<int>(src, 2);
   ASSERT_EQ(res.dims()[0], 2 * 3);
   ASSERT_EQ(res.dims()[1], 4 * 9);
-}
\ No newline at end of file
+}
diff --git a/paddle/function/neon/NeonDepthwiseConv.h b/paddle/function/neon/NeonDepthwiseConv.h
index aefeea78badbca3d0d09e292e4e1e148618f8ac6..33722d3cac61b62f5dce8f51105c1bf4e70c4a6c 100644
--- a/paddle/function/neon/NeonDepthwiseConv.h
+++ b/paddle/function/neon/NeonDepthwiseConv.h
@@ -594,7 +594,7 @@ struct StridePadding {
         float32x4_t s1 = vdupq_n_f32(0.f);
         for (int s = 0; s < step; s++) {
           float32x4_t s0 = vld1q_f32(input);
-          float32x4x2_t v = {s0, s1};
+          float32x4x2_t v = {{s0, s1}};
           vst2q_f32(inputPadding, v);
           input += 4;
           inputPadding += 8;
diff --git a/paddle/gserver/gradientmachines/RecurrentGradientMachine.cpp b/paddle/gserver/gradientmachines/RecurrentGradientMachine.cpp
index 9f29b97466910f1daf88e3ca86f92d10661462c5..b71431b90747e950a9099dafeb6403f2274eb130 100644
--- a/paddle/gserver/gradientmachines/RecurrentGradientMachine.cpp
+++ b/paddle/gserver/gradientmachines/RecurrentGradientMachine.cpp
@@ -18,6 +18,7 @@ limitations under the License. */
 #include <cmath>
 #include <functional>
 #include <limits>
+#include <memory>
 #include "NeuralNetwork.h"
 #include "paddle/gserver/layers/AgentLayer.h"
 #include "paddle/utils/Flags.h"
@@ -429,7 +430,11 @@ void RecurrentGradientMachine::reorganizeInput(PassType passType) {
   }
 
   {
-    AsyncGpuBlock asyncGpuBlock;
+    std::unique_ptr<AsyncGpuBlock> asyncBlock;
+
+    if (useGpu_) {
+      asyncBlock.reset(new AsyncGpuBlock());
+    }
 
     // inFrameLine select rows in real layer one time
     for (size_t i = 0; i < inFrameLines_.size(); i++) {
diff --git a/paddle/gserver/layers/ExpandConvBaseLayer.cpp b/paddle/gserver/layers/ExpandConvBaseLayer.cpp
deleted file mode 100644
index 2b7bef0a757d7c706be3815c539b036b094596cf..0000000000000000000000000000000000000000
--- a/paddle/gserver/layers/ExpandConvBaseLayer.cpp
+++ /dev/null
@@ -1,124 +0,0 @@
-/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve.
-
-Licensed under the Apache License, Version 2.0 (the "License");
-you may not use this file except in compliance with the License.
-You may obtain a copy of the License at
-
-    http://www.apache.org/licenses/LICENSE-2.0
-
-Unless required by applicable law or agreed to in writing, software
-distributed under the License is distributed on an "AS IS" BASIS,
-WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-See the License for the specific language governing permissions and
-limitations under the License. */
-
-#include "ExpandConvBaseLayer.h"
-
-#include "paddle/utils/Logging.h"
-namespace paddle {
-
-bool ExpandConvBaseLayer::init(const LayerMap &layerMap,
-                               const ParameterMap &parameterMap) {
-  /* Initialize the basic convolutional parent class */
-  ConvBaseLayer::init(layerMap, parameterMap);
-
-  int index = 0;
-  for (auto &inputConfig : config_.inputs()) {
-    const ConvConfig &conf = inputConfig.conv_conf();
-    /* Consistent caffe mode for multiple input */
-    caffeMode_ = conf.caffe_mode();
-
-    // create a new weight
-    size_t height, width;
-    height = filterPixels_[index] * filterChannels_[index];
-    width = (!isDeconv_) ? numFilters_ : channels_[index];
-    CHECK_EQ(parameters_[index]->getSize(), width * height);
-    Weight *w = new Weight(height, width, parameters_[index]);
-    weights_.emplace_back(w);
-    index++;
-  }
-  if (biasParameter_.get()) {
-    if (sharedBiases_) {
-      CHECK_EQ((size_t)numFilters_, biasParameter_->getSize());
-      biases_ =
-          std::unique_ptr<Weight>(new Weight(numFilters_, 1, biasParameter_));
-    } else {
-      biases_ =
-          std::unique_ptr<Weight>(new Weight(getSize(), 1, biasParameter_));
-    }
-  }
-  getOutputSize();
-
-  return true;
-}
-
-size_t ExpandConvBaseLayer::getOutputSize() {
-  CHECK_NE(inputLayers_.size(), 0UL);
-  size_t layerSize = ConvBaseLayer::calOutputSize();
-  return layerSize;
-}
-
-void ExpandConvBaseLayer::addSharedBias() {
-  size_t mapW = getOutputSize() / numFilters_;
-  size_t mapH = getOutputValue()->getElementCnt() / mapW;
-  MatrixPtr out =
-      Matrix::create(getOutputValue()->getData(), mapH, mapW, false, useGpu_);
-
-  Matrix::resizeOrCreate(transOutValue_, mapW, mapH, false, useGpu_);
-
-  out->transpose(transOutValue_, false);  // false means no memory allocation
-  transOutValue_->reshape(transOutValue_->getElementCnt() / numFilters_,
-                          numFilters_);
-
-  MatrixPtr bias = Matrix::create(biases_->getW()->getData(),
-                                  1,
-                                  biases_->getW()->getElementCnt(),
-                                  false,
-                                  useGpu_);
-  transOutValue_->addBias(*bias, 1.0f);
-
-  transOutValue_->reshape(mapW, mapH);
-  transOutValue_->transpose(out, false);  // false means no memory allocation
-
-  out->clear();
-  bias->clear();
-}
-
-void ExpandConvBaseLayer::addUnsharedBias() {
-  MatrixPtr outValue = getOutputValue();
-  MatrixPtr bias = Matrix::create(biases_->getW()->getData(),
-                                  1,
-                                  biases_->getW()->getElementCnt(),
-                                  false,
-                                  useGpu_);
-  outValue->addBias(*bias, 1.0f);
-}
-
-void ExpandConvBaseLayer::bpropSharedBias(MatrixPtr biases, MatrixPtr v) {
-  size_t mapW = getOutputSize() / numFilters_;
-  size_t mapH = v->getElementCnt() / mapW;
-  MatrixPtr vTmp = Matrix::create(v->getData(), mapH, mapW, false, useGpu_);
-
-  Matrix::resizeOrCreate(transOutValue_, mapW, mapH, false, useGpu_);
-
-  vTmp->transpose(transOutValue_, false);  // false means no memory allocation
-  transOutValue_->reshape(transOutValue_->getElementCnt() / numFilters_,
-                          numFilters_);
-  biases->collectBias(*transOutValue_, 1.0f);
-}
-
-void ExpandConvBaseLayer::bpropBiases(MatrixPtr v) {
-  MatrixPtr biases = Matrix::create(biases_->getWGrad()->getData(),
-                                    1,
-                                    biases_->getWGrad()->getElementCnt(),
-                                    false,
-                                    useGpu_);
-  if (sharedBiases_) {
-    bpropSharedBias(biases, v);
-  } else {
-    biases->collectBias(*v, 1.0f);
-  }
-  biases->clear();
-}
-
-}  // namespace paddle
diff --git a/paddle/gserver/layers/ExpandConvBaseLayer.h b/paddle/gserver/layers/ExpandConvBaseLayer.h
deleted file mode 100644
index 01c699d2344443a1887ec0b5005125f617cbe279..0000000000000000000000000000000000000000
--- a/paddle/gserver/layers/ExpandConvBaseLayer.h
+++ /dev/null
@@ -1,57 +0,0 @@
-/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve.
-
-Licensed under the Apache License, Version 2.0 (the "License");
-you may not use this file except in compliance with the License.
-You may obtain a copy of the License at
-
-    http://www.apache.org/licenses/LICENSE-2.0
-
-Unless required by applicable law or agreed to in writing, software
-distributed under the License is distributed on an "AS IS" BASIS,
-WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-See the License for the specific language governing permissions and
-limitations under the License. */
-
-#pragma once
-
-#include <vector>
-#include "ConvBaseLayer.h"
-#include "paddle/math/Matrix.h"
-
-namespace paddle {
-
-/**
- * @brief A subclass of ConvBaseLayer that is a superclass of both
- * ExpandConvLayer and ExpandConvTransLayer
- */
-class ExpandConvBaseLayer : public ConvBaseLayer {
-protected:
-  /// The transpose of output, which is an auxiliary matrix.
-  MatrixPtr transOutValue_;
-
-public:
-  explicit ExpandConvBaseLayer(const LayerConfig& config)
-      : ConvBaseLayer(config) {}
-
-  ~ExpandConvBaseLayer() {}
-
-  bool init(const LayerMap& layerMap,
-            const ParameterMap& parameterMap) override;
-
-  size_t getOutputSize();
-
-  /**
-   * Add shared bias.
-   */
-  void addSharedBias();
-
-  /**
-   * Add unshared bias.
-   */
-  void addUnsharedBias();
-
-  void bpropSharedBias(MatrixPtr biases, MatrixPtr v);
-  void bpropBiases(MatrixPtr v);
-};
-
-}  // namespace paddle
diff --git a/paddle/gserver/layers/ExpandConvLayer.cpp b/paddle/gserver/layers/ExpandConvLayer.cpp
index 20de475fc3f6b6f3c05ac26bea8363daff0cf110..48dfcb49a4c2c46891bb5236fc1f8e644c03f327 100644
--- a/paddle/gserver/layers/ExpandConvLayer.cpp
+++ b/paddle/gserver/layers/ExpandConvLayer.cpp
@@ -36,7 +36,36 @@ inline bool isDepthwiseConv(int channels, int groups) {
 bool ExpandConvLayer::init(const LayerMap &layerMap,
                            const ParameterMap &parameterMap) {
   /* Initialize the basic convolutional parent class */
-  ExpandConvBaseLayer::init(layerMap, parameterMap);
+  ConvBaseLayer::init(layerMap, parameterMap);
+
+  int index = 0;
+  for (auto &inputConfig : config_.inputs()) {
+    const ConvConfig &conf = inputConfig.conv_conf();
+    /* Consistent caffe mode for multiple input */
+    caffeMode_ = conf.caffe_mode();
+
+    // create a new weight
+    size_t height, width;
+    height = filterPixels_[index] * filterChannels_[index];
+    width = (!isDeconv_) ? numFilters_ : channels_[index];
+    CHECK_EQ(parameters_[index]->getSize(), width * height);
+    Weight *w = new Weight(height, width, parameters_[index]);
+    weights_.emplace_back(w);
+    index++;
+  }
+
+  if (biasParameter_.get()) {
+    if (sharedBiases_) {
+      CHECK_EQ((size_t)numFilters_, biasParameter_->getSize());
+      biases_ = std::unique_ptr<Weight>(
+          new Weight(1, numFilters_, biasParameter_, 0));
+    } else {
+      biases_ =
+          std::unique_ptr<Weight>(new Weight(1, getSize(), biasParameter_, 0));
+    }
+  }
+
+  getOutputSize();
 
   size_t numInputs = config_.inputs_size();
   inputShape_.resize(numInputs);
@@ -108,6 +137,12 @@ bool ExpandConvLayer::init(const LayerMap &layerMap,
   return true;
 }
 
+size_t ExpandConvLayer::getOutputSize() {
+  CHECK_NE(inputLayers_.size(), 0UL);
+  size_t layerSize = ConvBaseLayer::calOutputSize();
+  return layerSize;
+}
+
 // i is the index of input layers
 #define BACKWARD_INPUT(i, inputs, outputs) \
   backward_[2 * i]->calc(inputs, outputs)
@@ -155,11 +190,7 @@ void ExpandConvLayer::forward(PassType passType) {
 
   /* add the bias-vector */
   if (biases_.get()) {
-    if (sharedBiases_) {
-      addSharedBias();
-    } else {
-      addUnsharedBias();
-    }
+    output_.value->addBias(*biases_->getW(), 1.0, sharedBiases_);
   }
 
   /* activation */
@@ -171,7 +202,7 @@ void ExpandConvLayer::backward(const UpdateCallback &callback) {
 
   MatrixPtr outGrad = getOutputGrad();
   if (biases_ && biases_->getWGrad()) {
-    bpropBiases(outGrad);
+    biases_->getWGrad()->collectBias(*getOutputGrad(), 1, sharedBiases_);
     /* Increasing the number of gradient */
     biases_->getParameterPtr()->incUpdate(callback);
   }
diff --git a/paddle/gserver/layers/ExpandConvLayer.h b/paddle/gserver/layers/ExpandConvLayer.h
index a1f943d1521547af0f82cec7da8a4efe9037cd71..a0873de19253f2496bc0c2fba550b3199dfc7486 100644
--- a/paddle/gserver/layers/ExpandConvLayer.h
+++ b/paddle/gserver/layers/ExpandConvLayer.h
@@ -15,7 +15,7 @@ limitations under the License. */
 #pragma once
 
 #include <vector>
-#include "ExpandConvBaseLayer.h"
+#include "ConvBaseLayer.h"
 #include "paddle/math/Matrix.h"
 
 namespace paddle {
@@ -28,10 +28,9 @@ namespace paddle {
  * The config file api is img_conv_layer.
  */
 
-class ExpandConvLayer : public ExpandConvBaseLayer {
+class ExpandConvLayer : public ConvBaseLayer {
 public:
-  explicit ExpandConvLayer(const LayerConfig& config)
-      : ExpandConvBaseLayer(config) {}
+  explicit ExpandConvLayer(const LayerConfig& config) : ConvBaseLayer(config) {}
 
   ~ExpandConvLayer() {}
 
@@ -41,6 +40,8 @@ public:
   void forward(PassType passType) override;
   void backward(const UpdateCallback& callback) override;
 
+  size_t getOutputSize();
+
 protected:
   std::vector<TensorShape> inputShape_;
   std::vector<TensorShape> filterShape_;
diff --git a/paddle/gserver/layers/MKLDNNConvLayer.cpp b/paddle/gserver/layers/MKLDNNConvLayer.cpp
new file mode 100644
index 0000000000000000000000000000000000000000..9088744beebd25ac105737fe3b012de143c66a7c
--- /dev/null
+++ b/paddle/gserver/layers/MKLDNNConvLayer.cpp
@@ -0,0 +1,544 @@
+/* Copyright (c) 2017 PaddlePaddle Authors. All Rights Reserve.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+    http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License. */
+
+#include "MKLDNNConvLayer.h"
+#include "paddle/math/MathUtils.h"
+#include "paddle/utils/Logging.h"
+
+using namespace mkldnn;  // NOLINT
+typedef memory::format format;
+
+namespace paddle {
+
+REGISTER_LAYER(mkldnn_conv, MKLDNNConvLayer);
+
+bool MKLDNNConvLayer::init(const LayerMap& layerMap,
+                           const ParameterMap& parameterMap) {
+  if (!MKLDNNLayer::init(layerMap, parameterMap)) {
+    return false;
+  }
+  CHECK_EQ(inputLayers_.size(), 1) << "Only support one input layer yet";
+  CHECK_EQ(inputLayers_.size(), parameters_.size());
+  CHECK(config_.shared_biases()) << "Only support shared biases yet";
+
+  oc_ = config_.num_filters();
+  const ConvConfig& conf = config_.inputs(0).conv_conf();
+  ic_ = conf.channels();
+  fw_ = conf.filter_size();
+  fh_ = conf.filter_size_y();
+  pw_ = conf.padding();
+  ph_ = conf.padding_y();
+  dw_ = conf.dilation();
+  dh_ = conf.dilation_y();
+  sw_ = conf.stride();
+  sh_ = conf.stride_y();
+  gp_ = conf.groups();
+  oh_ = conf.output_y();
+  ow_ = conf.output_x();
+  ih_ = conf.img_size_y();
+  iw_ = conf.img_size();
+  caffeMode_ = conf.caffe_mode();
+  CHECK(caffeMode_) << "Only support caffe mode yet";
+  CHECK(dh_ == 1 && dw_ == 1) << "Only support dilation 1 yet";
+  // check group setting
+  CHECK_EQ((oc_ / gp_) * gp_, oc_) << "group is indivisible for oc";
+  CHECK_EQ((ic_ / gp_) * gp_, ic_) << "group is indivisible for ic";
+
+  // create weight
+  size_t height = oc_ / gp_;
+  size_t width = ic_ * fh_ * fw_;
+  CHECK_EQ(parameters_[0]->getSize(), height * width);
+  weight_ =
+      std::unique_ptr<Weight>(new Weight(height, width, parameters_[0], 0));
+
+  // create biases
+  if (biasParameter_.get() != NULL) {
+    biases_ = std::unique_ptr<Weight>(new Weight(1, oc_, biasParameter_));
+  }
+  return true;
+}
+
+void MKLDNNConvLayer::convertWeightsFromPaddle() {
+  if (hasInitedWgt_) {
+    return;
+  }
+
+  CHECK(wgtVal_) << "should have been initialized";
+  // the paddle weight format is oihw or goihw
+  auto targetDim = wgtVal_->getDims();
+  auto srcFmt = (gp_ == 1) ? memory::format::oihw : memory::format::goihw;
+  wgtVal_->reorderDataFrom(wgtVal_, srcFmt, targetDim);
+  hasInitedWgt_ = true;
+}
+
+void MKLDNNConvLayer::convertWeightsToPaddle() {
+  CHECK(wgtVal_) << "should have been initialized";
+  auto targetDim = wgtVal_->getDims();
+  auto dstFmt = (gp_ == 1) ? memory::format::oihw : memory::format::goihw;
+  wgtVal_->reorderDataTo(wgtVal_, dstFmt, targetDim);
+}
+
+void MKLDNNConvLayer::reshape(
+    int& bs, int& ic, int& ih, int& iw, int oc, int& oh, int& ow) {
+  reshapeInput(bs, ih, iw);
+
+  // cal output sizes
+  // oc can not be changed
+  int fh = (fh_ - 1) * dh_ + 1;
+  int fw = (fw_ - 1) * dw_ + 1;
+  oh = outputSize(ih, fh, ph_, sh_, caffeMode_);
+  ow = outputSize(iw, fw, pw_, sw_, caffeMode_);
+
+  reshapeOutput(oh, ow);
+  resizeOutput(bs, oc * oh * ow);
+
+  printSizeInfo();
+}
+
+void MKLDNNConvLayer::resetFwd(std::vector<primitive>& pipeline,
+                               MKLDNNMatrixPtr& in,
+                               MKLDNNMatrixPtr& wgt,
+                               MKLDNNMatrixPtr& bias,
+                               MKLDNNMatrixPtr& out) {
+  resetFwdPD(fwdPD_);
+
+  resetFwdBuffers(fwdPD_, in, wgt, bias, out);
+
+  resetFwdPipeline(pipeline, fwdPD_, in, wgt, bias, out);
+
+  printValueFormatFlow();
+}
+
+void MKLDNNConvLayer::resetBwd(std::vector<primitive>& pipeline,
+                               MKLDNNMatrixPtr& in,
+                               MKLDNNMatrixPtr& wgt,
+                               MKLDNNMatrixPtr& bias,
+                               MKLDNNMatrixPtr& out) {
+  std::shared_ptr<conv_bwdWgt::primitive_desc> bwdWgtPD;
+  std::shared_ptr<conv_bwdData::primitive_desc> bwdDataPD;
+
+  resetBwdWgtPD(bwdWgtPD);
+
+  resetBwdDataPD(bwdDataPD);
+
+  resetBwdBuffers(bwdWgtPD, bwdDataPD, in, wgt, bias, out);
+
+  resetBwdPipeline(pipeline, bwdWgtPD, bwdDataPD, in, wgt, bias, out);
+
+  printGradFormatFlow();
+}
+
+void MKLDNNConvLayer::updateInputData() {
+  cpuInVal_->setData(getInputValue(0, CPU_DEVICE)->getData());
+}
+
+void MKLDNNConvLayer::updateWeights(const UpdateCallback& callback) {
+  weight_->getParameterPtr()->incUpdate(callback);
+  if (biases_ && biases_->getWGrad()) {
+    biases_->getParameterPtr()->incUpdate(callback);
+  }
+}
+
+void MKLDNNConvLayer::loadConvSettings(memory::dims& wgt,
+                                       memory::dims& bias,
+                                       memory::dims& stride,
+                                       memory::dims& dilation,
+                                       memory::dims& padL,
+                                       memory::dims& padR) {
+  wgt = (gp_ == 1) ? memory::dims{oc_, ic_, fh_, fw_}
+                   : memory::dims{gp_, oc_ / gp_, ic_ / gp_, fh_, fw_};
+  bias = memory::dims{oc_};
+  stride = memory::dims{sh_, sw_};
+  padL = memory::dims{ph_, pw_};
+  padR = getPaddingR();
+  // note: mkldnn dilation start from 0
+  dilation = memory::dims{dh_ - 1, dw_ - 1};
+}
+
+void MKLDNNConvLayer::resetFwdPD(
+    std::shared_ptr<conv_fwd::primitive_desc>& pd) {
+  // dims for conv
+  memory::dims inDims = memory::dims{bs_, ic_, ih_, iw_};
+  memory::dims outDims = memory::dims{bs_, oc_, oh_, ow_};
+  memory::dims wgtDims, biasDims, strides, dilations, padL, padR;
+  loadConvSettings(wgtDims, biasDims, strides, dilations, padL, padR);
+
+  prop_kind pk = passType_ == PASS_TEST ? prop_kind::forward_scoring
+                                        : prop_kind::forward_training;
+  algorithm algo = algorithm::convolution_direct;
+  padding_kind padKind = padding_kind::zero;
+  conv_fwd::desc fwdDesc =
+      biases_ && biases_->getW()
+          ? conv_fwd::desc(pk,
+                           algo,
+                           MKLDNNMatrix::createMemoryDesc(inDims),
+                           MKLDNNMatrix::createMemoryDesc(wgtDims),
+                           MKLDNNMatrix::createMemoryDesc(biasDims),
+                           MKLDNNMatrix::createMemoryDesc(outDims),
+                           strides,
+                           dilations,
+                           padL,
+                           padR,
+                           padKind)
+          : conv_fwd::desc(pk,
+                           algo,
+                           MKLDNNMatrix::createMemoryDesc(inDims),
+                           MKLDNNMatrix::createMemoryDesc(wgtDims),
+                           MKLDNNMatrix::createMemoryDesc(outDims),
+                           strides,
+                           dilations,
+                           padL,
+                           padR,
+                           padKind);
+  pd.reset(new conv_fwd::primitive_desc(fwdDesc, engine_));
+}
+
+void MKLDNNConvLayer::resetFwdBuffers(
+    std::shared_ptr<conv_fwd::primitive_desc>& pd,
+    MKLDNNMatrixPtr& in,
+    MKLDNNMatrixPtr& wgt,
+    MKLDNNMatrixPtr& bias,
+    MKLDNNMatrixPtr& out) {
+  CHECK(pd);
+  resetInValue(pd, in);
+
+  resetWgtBiasValue(pd, wgt, bias);
+
+  resetOutValue(pd, out);
+}
+
+void MKLDNNConvLayer::resetFwdPipeline(
+    std::vector<primitive>& pipeline,
+    std::shared_ptr<conv_fwd::primitive_desc>& pd,
+    MKLDNNMatrixPtr& in,
+    MKLDNNMatrixPtr& wgt,
+    MKLDNNMatrixPtr& bias,
+    MKLDNNMatrixPtr& out) {
+  pipeline.clear();
+
+  if (cvtInVal_) {
+    pipeline.push_back(*cvtInVal_);
+  }
+
+  if (bias) {
+    fwd_.reset(new conv_fwd(*pd, *in, *wgt, *bias, *out));
+  } else {
+    fwd_.reset(new conv_fwd(*pd, *in, *wgt, *out));
+  }
+  pipeline.push_back(*fwd_);
+
+  if (cvtOutVal_) {
+    pipeline.push_back(*cvtOutVal_);
+  }
+}
+
+void MKLDNNConvLayer::resetInValue(
+    std::shared_ptr<conv_fwd::primitive_desc>& pd, MKLDNNMatrixPtr& in) {
+  const MatrixPtr& inMat = inputLayers_[0]->getOutput().value;
+  in = MKLDNNMatrix::create(inMat, pd->src_primitive_desc());
+
+  // create buffer and reorder if input value do not match
+  cpuInVal_ = nullptr;
+  cvtInVal_ = nullptr;
+  if (inputIsOnlyMKLDNN()) {
+    MKLDNNMatrixPtr dnnIn = std::dynamic_pointer_cast<MKLDNNMatrix>(inMat);
+    CHECK(dnnIn) << "Input should be MKLDNNMatrix";
+    if (dnnIn->getPrimitiveDesc() != in->getPrimitiveDesc()) {
+      CHECK_EQ(dnnIn->getFormat(), format::nc);
+      CHECK(ih_ == 1 && iw_ == 1) << "when input is nc format";
+      // create a new one with nchw format and same data
+      memory::dims inDims = memory::dims{bs_, ic_, 1, 1};
+      dnnIn = MKLDNNMatrix::create(inMat, inDims, format::nchw, engine_);
+      CHECK(dnnIn->getPrimitiveDesc() == in->getPrimitiveDesc());
+    }
+    in = dnnIn;
+  } else {
+    const MatrixPtr& cpuIn = getInputValue(0, CPU_DEVICE);
+    memory::dims inDims = memory::dims{bs_, ic_, ih_, iw_};
+    cpuInVal_ = MKLDNNMatrix::create(cpuIn, inDims, format::nchw, engine_);
+    if (cpuInVal_->getPrimitiveDesc() != in->getPrimitiveDesc()) {
+      // create new mkldnn matrix
+      in = MKLDNNMatrix::create(nullptr, pd->src_primitive_desc());
+      cvtInVal_ = MKLDNNMatrix::createReorder(cpuInVal_, in);
+      CHECK(cvtInVal_) << "should not be emptry";
+    } else {
+      in = cpuInVal_;
+    }
+  }
+}
+
+void MKLDNNConvLayer::resetWgtBiasValue(
+    std::shared_ptr<conv_fwd::primitive_desc>& pd,
+    MKLDNNMatrixPtr& wgt,
+    MKLDNNMatrixPtr& bias) {
+  wgt = MKLDNNMatrix::create(weight_->getW(), pd->weights_primitive_desc());
+  VLOG(MKLDNN_FMTS) << "Weight value format: " << wgt->getFormat();
+
+  bias = (biases_ && biases_->getW())
+             ? MKLDNNMatrix::create(biases_->getW(), pd->bias_primitive_desc())
+             : nullptr;
+}
+
+void MKLDNNConvLayer::resetOutValue(
+    std::shared_ptr<conv_fwd::primitive_desc>& pd, MKLDNNMatrixPtr& out) {
+  out = MKLDNNMatrix::create(output_.value, pd->dst_primitive_desc());
+
+  // change original output value from cpu matrix to mkldnn matrix
+  output_.value = std::dynamic_pointer_cast<Matrix>(out);
+
+  // create reorder if output value has cpu device and pd do not match
+  cpuOutVal_ = nullptr;
+  cpuOutVal_ = nullptr;
+  if (!outputIsOnlyMKLDNN()) {
+    const MatrixPtr& cpuOut = getOutput(CPU_DEVICE).value;
+    memory::dims outDims = memory::dims{bs_, oc_, oh_, ow_};
+    cpuOutVal_ = MKLDNNMatrix::create(cpuOut, outDims, format::nchw, engine_);
+    if (cpuOutVal_->getPrimitiveDesc() != out->getPrimitiveDesc()) {
+      cvtOutVal_ = MKLDNNMatrix::createReorder(out, cpuOutVal_);
+      CHECK(cvtOutVal_) << "should not be emptry";
+    } else {
+      // CPU output share the same data of MKLDNN output
+      cpuOut->setData(out->getData());
+      cpuOutVal_ = out;
+    }
+  }
+}
+
+void MKLDNNConvLayer::resetBwdWgtPD(
+    std::shared_ptr<conv_bwdWgt::primitive_desc>& pd) {
+  memory::dims wgtDims, biasDims, strides, dilations, padL, padR;
+  loadConvSettings(wgtDims, biasDims, strides, dilations, padL, padR);
+
+  // create backward weight using input, output and weight value memory desc
+  CHECK(inVal_) << "Should have input value";
+  CHECK(outVal_) << "Should have output value";
+  CHECK(wgtVal_) << "Should have weight value";
+  algorithm algo = algorithm::convolution_direct;
+  padding_kind padKind = padding_kind::zero;
+  auto bwdWgtDesc = biasVal_ != nullptr
+                        ? conv_bwdWgt::desc(algo,
+                                            inVal_->getMemoryDesc(),
+                                            wgtVal_->getMemoryDesc(),
+                                            biasVal_->getMemoryDesc(),
+                                            outVal_->getMemoryDesc(),
+                                            strides,
+                                            padL,
+                                            padR,
+                                            padKind)
+                        : conv_bwdWgt::desc(algo,
+                                            inVal_->getMemoryDesc(),
+                                            wgtVal_->getMemoryDesc(),
+                                            outVal_->getMemoryDesc(),
+                                            strides,
+                                            padL,
+                                            padR,
+                                            padKind);
+  pd.reset(new conv_bwdWgt::primitive_desc(bwdWgtDesc, engine_, *fwdPD_));
+  CHECK(pd->src_primitive_desc() == inVal_->getPrimitiveDesc())
+      << "primitive desc of in value should equal";
+  CHECK(pd->diff_dst_primitive_desc() == outVal_->getPrimitiveDesc())
+      << "primitive desc of out grad should equal the out value";
+  CHECK(pd->diff_weights_primitive_desc() == wgtVal_->getPrimitiveDesc())
+      << "primitive desc of weight grad should equal the weight value";
+}
+
+void MKLDNNConvLayer::resetBwdDataPD(
+    std::shared_ptr<conv_bwdData::primitive_desc>& pd) {
+  pd = nullptr;
+  if (inputLayers_[0]->getOutput().grad == nullptr) {
+    return;
+  }
+
+  memory::dims wgtDims, biasDims, strides, dilations, padL, padR;
+  loadConvSettings(wgtDims, biasDims, strides, dilations, padL, padR);
+  CHECK(inVal_) << "Should have input value";
+  CHECK(outVal_) << "Should have output value";
+  // create backward data using input and output value memory desc
+  // but using weight memory desc with any format
+  auto bwdDataDesc = conv_bwdData::desc(algorithm::convolution_direct,
+                                        inVal_->getMemoryDesc(),
+                                        MKLDNNMatrix::createMemoryDesc(wgtDims),
+                                        outVal_->getMemoryDesc(),
+                                        strides,
+                                        padL,
+                                        padR,
+                                        padding_kind::zero);
+  pd.reset(new conv_bwdData::primitive_desc(bwdDataDesc, engine_, *fwdPD_));
+  CHECK(pd->diff_src_primitive_desc() == inVal_->getPrimitiveDesc())
+      << "primitive desc of in grad should equal the in value";
+  CHECK(pd->diff_dst_primitive_desc() == outVal_->getPrimitiveDesc())
+      << "primitive desc of out grad should equal";
+}
+
+void MKLDNNConvLayer::resetBwdBuffers(
+    std::shared_ptr<conv_bwdWgt::primitive_desc>& wgtPD,
+    std::shared_ptr<conv_bwdData::primitive_desc>& dataPD,
+    MKLDNNMatrixPtr& in,
+    MKLDNNMatrixPtr& wgt,
+    MKLDNNMatrixPtr& bias,
+    MKLDNNMatrixPtr& out) {
+  CHECK(wgtPD);
+  resetOutGrad(wgtPD, out);
+
+  resetWgtBiasGrad(wgtPD, wgt, bias);
+
+  resetInGrad(dataPD, in);
+
+  resetWgtValBwdData(dataPD, wgtValBwdData_);
+}
+
+void MKLDNNConvLayer::resetBwdPipeline(
+    std::vector<primitive>& pipeline,
+    std::shared_ptr<conv_bwdWgt::primitive_desc>& wgtPD,
+    std::shared_ptr<conv_bwdData::primitive_desc>& dataPD,
+    MKLDNNMatrixPtr& in,
+    MKLDNNMatrixPtr& wgt,
+    MKLDNNMatrixPtr& bias,
+    MKLDNNMatrixPtr& out) {
+  pipeline.clear();
+
+  if (cvtOutGrad_) {
+    pipeline.push_back(*cvtOutGrad_);
+  }
+
+  // add bwdWgt handle
+  if (bias) {
+    bwdWgt_.reset(new conv_bwdWgt(*wgtPD, *inVal_, *out, *wgt, *bias));
+  } else {
+    bwdWgt_.reset(new conv_bwdWgt(*wgtPD, *inVal_, *out, *wgt));
+  }
+  pipeline.push_back(*bwdWgt_);
+
+  if (dataPD == nullptr) {
+    return;
+  }
+
+  if (cvtWgtVal_) {
+    pipeline.push_back(*cvtWgtVal_);
+  }
+
+  // add bwdData handle
+  CHECK(wgtValBwdData_) << "Should have weight memory";
+  bwdData_.reset(new conv_bwdData(*dataPD, *out, *wgtValBwdData_, *in));
+  pipeline.push_back(*bwdData_);
+
+  if (cvtInGrad_) {
+    pipeline.push_back(*cvtInGrad_);
+  }
+}
+
+void MKLDNNConvLayer::resetOutGrad(
+    std::shared_ptr<conv_bwdWgt::primitive_desc>& wgtPD, MKLDNNMatrixPtr& out) {
+  const MatrixPtr& outMat = output_.grad;
+  out = MKLDNNMatrix::create(outMat, wgtPD->diff_dst_primitive_desc());
+  CHECK(outVal_ != nullptr &&
+        out->getPrimitiveDesc() == outVal_->getPrimitiveDesc())
+      << "primitive desc of out grad and value should be equal";
+
+  // TODO(TJ): merge outgrad
+  // create reorder if has output grad does not match
+  cpuOutGrad_ = nullptr;
+  cvtOutGrad_ = nullptr;
+  if (!outputIsOnlyMKLDNN()) {
+    const MatrixPtr& cpuOut = getOutput(CPU_DEVICE).grad;
+    // same PrimitiveDesc with cpuInVal_
+    CHECK(cpuOutVal_);
+    cpuOutGrad_ = MKLDNNMatrix::create(cpuOut, cpuOutVal_->getPrimitiveDesc());
+    if (cpuOutGrad_->getPrimitiveDesc() == out->getPrimitiveDesc()) {
+      outMat->setData(cpuOut->getData());
+      out = cpuOutGrad_;
+    } else {
+      cvtOutGrad_ = MKLDNNMatrix::createReorder(cpuOutGrad_, out);
+      CHECK(cvtOutGrad_);
+    }
+  }
+}
+
+void MKLDNNConvLayer::resetWgtBiasGrad(
+    std::shared_ptr<conv_bwdWgt::primitive_desc>& wgtPD,
+    MKLDNNMatrixPtr& wgt,
+    MKLDNNMatrixPtr& bias) {
+  wgt = MKLDNNMatrix::create(weight_->getWGrad(),
+                             wgtPD->diff_weights_primitive_desc());
+  CHECK(nullptr != wgtVal_ &&
+        wgt->getPrimitiveDesc() == wgtVal_->getPrimitiveDesc())
+      << "primitive desc of weight grad and value should be equal";
+  VLOG(MKLDNN_FMTS) << "weight grad format: " << wgt->getFormat();
+
+  bias = nullptr;
+  if (biasVal_ == nullptr) {
+    return;
+  }
+  bias = MKLDNNMatrix::create(biases_->getWGrad(),
+                              wgtPD->diff_bias_primitive_desc());
+  CHECK(bias->getPrimitiveDesc() == biasVal_->getPrimitiveDesc())
+      << "primitive desc of bias grad should equal the bias value";
+}
+
+void MKLDNNConvLayer::resetInGrad(
+    std::shared_ptr<conv_bwdData::primitive_desc>& dataPD,
+    MKLDNNMatrixPtr& in) {
+  if (dataPD == nullptr) {
+    return;
+  }
+
+  // TODO(TJ): use outputMaps_ ways to get the inGrad_ when merge outgrad done
+  in = MKLDNNMatrix::create(inputLayers_[0]->getOutput().grad,
+                            dataPD->diff_src_primitive_desc());
+  CHECK(nullptr != inVal_ &&
+        in->getPrimitiveDesc() == inVal_->getPrimitiveDesc())
+      << "primitive desc of input grad and value should be equal";
+
+  // create reorder if has output grad does not match
+  cpuInGrad_ = nullptr;
+  cvtInGrad_ = nullptr;
+  if (!inputIsOnlyMKLDNN()) {
+    const MatrixPtr& cpuIn = getInputGrad(0, CPU_DEVICE);
+    // same PrimitiveDesc with cpuInVal_
+    CHECK(cpuInVal_);
+    cpuInGrad_ = MKLDNNMatrix::create(cpuIn, cpuInVal_->getPrimitiveDesc());
+    if (cpuInGrad_->getPrimitiveDesc() != in->getPrimitiveDesc()) {
+      const MatrixPtr& dnnIn = getInputGrad(0, MKLDNN_DEVICE);
+      in = MKLDNNMatrix::create(dnnIn, in->getPrimitiveDesc());
+      cvtInGrad_ = MKLDNNMatrix::createReorder(in, cpuInGrad_);
+      CHECK(cvtInGrad_);
+    } else {
+      in = cpuInGrad_;
+    }
+  }
+}
+
+void MKLDNNConvLayer::resetWgtValBwdData(
+    std::shared_ptr<conv_bwdData::primitive_desc>& dataPD,
+    MKLDNNMatrixPtr& wgt) {
+  if (dataPD == nullptr) {
+    return;
+  }
+
+  // create new weight value for backward data, and create reorder if necessary
+  // since the primitive_desc would be different with wgtVal_
+  CHECK(wgtVal_) << "should have weight value";
+  if (dataPD->weights_primitive_desc() != wgtVal_->getPrimitiveDesc()) {
+    wgtValBwdData_ =
+        MKLDNNMatrix::create(nullptr, dataPD->weights_primitive_desc());
+    cvtWgtVal_ = MKLDNNMatrix::createReorder(wgtVal_, wgtValBwdData_);
+    CHECK(cvtWgtVal_);
+  } else {
+    wgtValBwdData_ = wgtVal_;
+  }
+  VLOG(MKLDNN_FMTS) << "weight value format for backward data"
+                    << wgtValBwdData_->getFormat();
+}
+
+}  // namespace paddle
diff --git a/paddle/gserver/layers/MKLDNNConvLayer.h b/paddle/gserver/layers/MKLDNNConvLayer.h
new file mode 100644
index 0000000000000000000000000000000000000000..f84f2f737c47a1b8adc2b83360a0396ffbc6ae24
--- /dev/null
+++ b/paddle/gserver/layers/MKLDNNConvLayer.h
@@ -0,0 +1,253 @@
+/* Copyright (c) 2017 PaddlePaddle Authors. All Rights Reserve.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+    http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License. */
+
+#pragma once
+
+#include "MKLDNNLayer.h"
+#include "mkldnn.hpp"
+
+namespace paddle {
+typedef mkldnn::convolution_forward conv_fwd;
+typedef mkldnn::convolution_backward_weights conv_bwdWgt;
+typedef mkldnn::convolution_backward_data conv_bwdData;
+
+/**
+ * @brief A subclass of MKLDNNLayer conv layer.
+ *
+ * The config file api is mkldnn_conv
+ */
+class MKLDNNConvLayer : public MKLDNNLayer {
+protected:
+  // padding height and width
+  int ph_, pw_;
+  // stride height and width
+  int sh_, sw_;
+  // dilation height and width
+  int dh_, dw_;
+  // filter(kenerl) height and width
+  int fh_, fw_;
+  // group number
+  int gp_;
+
+  // in resetBwdData, the format of wgtValBwdData_ is different with wgtVal_
+  MKLDNNMatrixPtr wgtValBwdData_;
+  // convert handle from wgtVal_ to wgtValBwdData_
+  std::shared_ptr<mkldnn::reorder> cvtWgtVal_;
+
+  // save forward primitive_desc, which can be used backward
+  std::shared_ptr<conv_fwd::primitive_desc> fwdPD_;
+
+  // MKLDNNMatrixPtr which should be created from CPU Device
+  MKLDNNMatrixPtr cpuInVal_;
+  MKLDNNMatrixPtr cpuInGrad_;
+  MKLDNNMatrixPtr cpuOutVal_;
+  MKLDNNMatrixPtr cpuOutGrad_;
+  // convert handle between CPU device and MKLDNN device
+  std::shared_ptr<mkldnn::reorder> cvtInVal_;
+  std::shared_ptr<mkldnn::reorder> cvtInGrad_;
+  std::shared_ptr<mkldnn::reorder> cvtOutVal_;
+  std::shared_ptr<mkldnn::reorder> cvtOutGrad_;
+
+  // whether the weight has been init
+  bool hasInitedWgt_;
+
+  // true by default, which impact the calculation of output image size.
+  // details can refer to mathUtil.h
+  bool caffeMode_;
+
+  // weight and bias
+  std::unique_ptr<Weight> weight_;
+  std::unique_ptr<Weight> biases_;
+
+public:
+  explicit MKLDNNConvLayer(const LayerConfig& config)
+      : MKLDNNLayer(config), hasInitedWgt_(false), caffeMode_(true) {}
+
+  ~MKLDNNConvLayer() {}
+
+  bool init(const LayerMap& layerMap,
+            const ParameterMap& parameterMap) override;
+
+  void reshape(
+      int& bs, int& ic, int& ih, int& iw, int oc, int& oh, int& ow) override;
+
+  void resetFwd(std::vector<mkldnn::primitive>& pipeline,
+                MKLDNNMatrixPtr& in,
+                MKLDNNMatrixPtr& wgt,
+                MKLDNNMatrixPtr& bias,
+                MKLDNNMatrixPtr& out) override;
+
+  void resetBwd(std::vector<mkldnn::primitive>& pipeline,
+                MKLDNNMatrixPtr& in,
+                MKLDNNMatrixPtr& wgt,
+                MKLDNNMatrixPtr& bias,
+                MKLDNNMatrixPtr& out) override;
+
+  void updateInputData() override;
+
+  void updateWeights(const UpdateCallback& callback) override;
+
+  void convertWeightsFromPaddle() override;
+
+  void convertWeightsToPaddle() override;
+
+  void printSizeInfo() override {
+    MKLDNNLayer::printSizeInfo();
+    VLOG(MKLDNN_SIZES) << getName() << ": fh: " << fh_ << ", fw: " << fw_
+                       << ": ph: " << ph_ << ", pw: " << pw_ << ", sh: " << sh_
+                       << ", sw: " << sw_ << ", dh: " << dh_ << ", dw: " << dw_;
+  }
+
+  void printValueFormatFlow() override {
+    if (cpuInVal_) {
+      VLOG(MKLDNN_FMTS) << cpuInVal_->getFormat() << " >>>";
+    }
+    MKLDNNLayer::printValueFormatFlow();
+    if (cpuOutVal_) {
+      VLOG(MKLDNN_FMTS) << " >>> " << cpuOutVal_->getFormat();
+    }
+  }
+
+  void printGradFormatFlow() override {
+    if (cpuInGrad_) {
+      VLOG(MKLDNN_FMTS) << cpuInGrad_->getFormat() << " <<<";
+    }
+    MKLDNNLayer::printGradFormatFlow();
+    if (cpuOutGrad_) {
+      VLOG(MKLDNN_FMTS) << " <<< " << cpuOutGrad_->getFormat();
+    }
+  }
+
+protected:
+  /**
+   * load the dims settings of this conv
+   */
+  void loadConvSettings(mkldnn::memory::dims& wgt,
+                        mkldnn::memory::dims& bias,
+                        mkldnn::memory::dims& stride,
+                        mkldnn::memory::dims& dilation,
+                        mkldnn::memory::dims& padL,
+                        mkldnn::memory::dims& padR);
+
+  /**
+   * reset the forward primitive descriptor.
+   */
+  void resetFwdPD(std::shared_ptr<conv_fwd::primitive_desc>& pd);
+  /**
+   * reset the MKLDNNMatrix buffers used in forward.
+   */
+  void resetFwdBuffers(std::shared_ptr<conv_fwd::primitive_desc>& pd,
+                       MKLDNNMatrixPtr& in,
+                       MKLDNNMatrixPtr& wgt,
+                       MKLDNNMatrixPtr& bias,
+                       MKLDNNMatrixPtr& out);
+  /**
+   * reset the forward pipeline.
+   */
+  void resetFwdPipeline(std::vector<mkldnn::primitive>& pipeline,
+                        std::shared_ptr<conv_fwd::primitive_desc>& pd,
+                        MKLDNNMatrixPtr& in,
+                        MKLDNNMatrixPtr& wgt,
+                        MKLDNNMatrixPtr& bias,
+                        MKLDNNMatrixPtr& out);
+
+  /**
+   * reset MKLDNNMatrix of input value
+   */
+  void resetInValue(std::shared_ptr<conv_fwd::primitive_desc>& pd,
+                    MKLDNNMatrixPtr& in);
+  /**
+   * reset MKLDNNMatrix of weight and bias value
+   */
+  void resetWgtBiasValue(std::shared_ptr<conv_fwd::primitive_desc>& pd,
+                         MKLDNNMatrixPtr& wgt,
+                         MKLDNNMatrixPtr& bias);
+  /**
+   * reset MKLDNNMatrix of output value
+   */
+  void resetOutValue(std::shared_ptr<conv_fwd::primitive_desc>& pd,
+                     MKLDNNMatrixPtr& out);
+
+  /**
+   * reset the backward weight primitive descriptor.
+   */
+  void resetBwdWgtPD(std::shared_ptr<conv_bwdWgt::primitive_desc>& pd);
+  /**
+   * reset the backward data primitive descriptor.
+   */
+  void resetBwdDataPD(std::shared_ptr<conv_bwdData::primitive_desc>& pd);
+  /**
+   * reset the MKLDNNMatrix buffers used in backward.
+   */
+  void resetBwdBuffers(std::shared_ptr<conv_bwdWgt::primitive_desc>& wgtPD,
+                       std::shared_ptr<conv_bwdData::primitive_desc>& dataPD,
+                       MKLDNNMatrixPtr& in,
+                       MKLDNNMatrixPtr& wgt,
+                       MKLDNNMatrixPtr& bias,
+                       MKLDNNMatrixPtr& out);
+  /**
+   * reset the backward pipeline.
+   */
+  void resetBwdPipeline(std::vector<mkldnn::primitive>& pipeline,
+                        std::shared_ptr<conv_bwdWgt::primitive_desc>& wgtPD,
+                        std::shared_ptr<conv_bwdData::primitive_desc>& dataPD,
+                        MKLDNNMatrixPtr& in,
+                        MKLDNNMatrixPtr& wgt,
+                        MKLDNNMatrixPtr& bias,
+                        MKLDNNMatrixPtr& out);
+
+  /**
+   * reset MKLDNNMatrix of output grad
+   */
+  void resetOutGrad(std::shared_ptr<conv_bwdWgt::primitive_desc>& wgtPD,
+                    MKLDNNMatrixPtr& out);
+  /**
+   * reset MKLDNNMatrix of weight and bias grad
+   */
+  void resetWgtBiasGrad(std::shared_ptr<conv_bwdWgt::primitive_desc>& wgtPD,
+                        MKLDNNMatrixPtr& wgt,
+                        MKLDNNMatrixPtr& bias);
+  /**
+   * reset MKLDNNMatrix of input grad
+   */
+  void resetInGrad(std::shared_ptr<conv_bwdData::primitive_desc>& dataPD,
+                   MKLDNNMatrixPtr& in);
+  /**
+   * reset MKLDNNMatrix of weight value for backward data
+   * since the primitive_desc would be different with wgtVal_
+   */
+  void resetWgtValBwdData(std::shared_ptr<conv_bwdData::primitive_desc>& dataPD,
+                          MKLDNNMatrixPtr& wgt);
+
+  /**
+   * get padding_r according to
+   * https://github.com/01org/mkl-dnn/blob/master/tests/gtests/
+   * test_convolution_forward_common.hpp
+   * @note: mkldnn dilation start from 0 while paddle start from 1
+   */
+  mkldnn::memory::dims getPaddingR() const {
+    mkldnn::memory::dims padR = {ph_, pw_};
+    for (int i = 0; i < 2; ++i) {
+      if ((ih_ - ((fh_ - 1) * dh_ + 1) + ph_ + padR[0]) / sh_ + 1 != oh_) {
+        ++padR[0];
+      }
+      if ((iw_ - ((fw_ - 1) * dw_ + 1) + pw_ + padR[1]) / sw_ + 1 != ow_) {
+        ++padR[1];
+      }
+    }
+    return padR;
+  }
+};
+
+}  // namespace paddle
diff --git a/paddle/gserver/layers/MKLDNNFcLayer.cpp b/paddle/gserver/layers/MKLDNNFcLayer.cpp
index f70343251ad4fbb99f9614618f6d1bff1174f15e..f60e221a6ec2ff513789a24e9f59bb25aef437b5 100644
--- a/paddle/gserver/layers/MKLDNNFcLayer.cpp
+++ b/paddle/gserver/layers/MKLDNNFcLayer.cpp
@@ -17,9 +17,6 @@ limitations under the License. */
 
 using namespace mkldnn;  // NOLINT
 typedef memory::format format;
-typedef inner_product_forward fc_fwd;
-typedef inner_product_backward_weights fc_bwdWgt;
-typedef inner_product_backward_data fc_bwdData;
 
 namespace paddle {
 
@@ -93,35 +90,88 @@ void MKLDNNFcLayer::reshape(
   printSizeInfo();
 }
 
-void MKLDNNFcLayer::resetFwd(std::vector<mkldnn::primitive>& pipeline,
+void MKLDNNFcLayer::resetFwd(std::vector<primitive>& pipeline,
                              MKLDNNMatrixPtr& in,
                              MKLDNNMatrixPtr& wgt,
                              MKLDNNMatrixPtr& bias,
                              MKLDNNMatrixPtr& out) {
-  pipeline.clear();
-  bool hasBias = biases_ && biases_->getW();
-  const MatrixPtr& wgtVal = weight_->getW();
-  const MatrixPtr& biasVal = hasBias ? biases_->getW() : nullptr;
-  const MatrixPtr& outVal = output_.value;
+  resetFwdBuffers(in, wgt, bias, out);
+
+  resetFwdPD(fwdPD_, in, wgt, bias, out);
+
+  resetFwdPipeline(pipeline, fwdPD_, in, wgt, bias, out);
+
+  printValueFormatFlow();
+}
+
+void MKLDNNFcLayer::resetBwd(std::vector<primitive>& pipeline,
+                             MKLDNNMatrixPtr& in,
+                             MKLDNNMatrixPtr& wgt,
+                             MKLDNNMatrixPtr& bias,
+                             MKLDNNMatrixPtr& out) {
+  std::shared_ptr<fc_bwdWgt::primitive_desc> bwdWgtPD;
+  std::shared_ptr<fc_bwdData::primitive_desc> bwdDataPD;
+
+  resetBwdBuffers(in, wgt, bias, out);
+
+  resetBwdWgtPD(bwdWgtPD, wgt, bias, out);
+
+  resetBwdDataPD(bwdDataPD, in, out);
+
+  resetBwdPipeline(pipeline, bwdWgtPD, bwdDataPD, in, wgt, bias, out);
+
+  printGradFormatFlow();
+}
+
+void MKLDNNFcLayer::updateInputData() {
+  inVal_->setData(getInputValue(0, CPU_DEVICE)->getData());
+}
 
+void MKLDNNFcLayer::updateWeights(const UpdateCallback& callback) {
+  weight_->getParameterPtr()->incUpdate(callback);
+  if (biases_ && biases_->getWGrad()) {
+    biases_->getParameterPtr()->incUpdate(callback);
+  }
+}
+
+void MKLDNNFcLayer::resetFwdBuffers(MKLDNNMatrixPtr& in,
+                                    MKLDNNMatrixPtr& wgt,
+                                    MKLDNNMatrixPtr& bias,
+                                    MKLDNNMatrixPtr& out) {
+  resetInValue(in);
+
+  resetWgtBiasValue(wgt, bias);
+
+  resetOutValue(out);
+}
+
+void MKLDNNFcLayer::resetInValue(MKLDNNMatrixPtr& in) {
   if (inputIsOnlyMKLDNN()) {
-    const MatrixPtr& inVal = getInputValue(0);
-    in = std::dynamic_pointer_cast<MKLDNNMatrix>(inVal);
+    const MatrixPtr& dnnIn = getInputValue(0);
+    in = std::dynamic_pointer_cast<MKLDNNMatrix>(dnnIn);
     CHECK(in) << "Input should be MKLDNNMatrix";
   } else {
     CHECK_EQ(getPrev(0)->getDeviceId(), CPU_DEVICE) << "Only support CPU yet";
-    const MatrixPtr& inVal = getInputValue(0, CPU_DEVICE);
+    const MatrixPtr& cpuIn = getInputValue(0, CPU_DEVICE);
     in = MKLDNNMatrix::create(
-        inVal, memory::dims{bs_, ic_, ih_, iw_}, format::nchw, engine_);
+        cpuIn, {bs_, ic_, ih_, iw_}, format::nchw, engine_);
   }
   in->downSpatial();
+}
+
+void MKLDNNFcLayer::resetWgtBiasValue(MKLDNNMatrixPtr& wgt,
+                                      MKLDNNMatrixPtr& bias) {
   wgt = MKLDNNMatrix::create(
-      wgtVal, memory::dims{oc_, ic_, ih_, iw_}, format::oihw, engine_);
+      weight_->getW(), {oc_, ic_, ih_, iw_}, format::oihw, engine_);
   wgt->downSpatial();
-  bias = hasBias ? MKLDNNMatrix::create(biasVal, {oc_}, format::x, engine_)
-                 : nullptr;
-  out = MKLDNNMatrix::create(outVal, {bs_, oc_}, format::nc, engine_);
 
+  bias = (biases_ && biases_->getW())
+             ? MKLDNNMatrix::create(biases_->getW(), {oc_}, format::x, engine_)
+             : nullptr;
+}
+
+void MKLDNNFcLayer::resetOutValue(MKLDNNMatrixPtr& out) {
+  out = MKLDNNMatrix::create(output_.value, {bs_, oc_}, format::nc, engine_);
   // change original output value to mkldnn output value
   output_.value = std::dynamic_pointer_cast<Matrix>(out);
   if (!outputIsOnlyMKLDNN()) {
@@ -129,46 +179,59 @@ void MKLDNNFcLayer::resetFwd(std::vector<mkldnn::primitive>& pipeline,
     // just share point
     getOutput(CPU_DEVICE).value->setData(output_.value->getData());
   }
+}
 
-  // create forward handle
+void MKLDNNFcLayer::resetFwdPD(std::shared_ptr<fc_fwd::primitive_desc>& pd,
+                               MKLDNNMatrixPtr in,
+                               MKLDNNMatrixPtr wgt,
+                               MKLDNNMatrixPtr bias,
+                               MKLDNNMatrixPtr out) {
+  CHECK(in);
+  CHECK(wgt);
+  CHECK(out);
   prop_kind pk = prop_kind::forward;
-  fc_fwd::desc fwdDesc = hasBias ? fc_fwd::desc(pk,
-                                                in->getMemoryDesc(),
-                                                wgt->getMemoryDesc(),
-                                                bias->getMemoryDesc(),
-                                                out->getMemoryDesc())
-                                 : fc_fwd::desc(pk,
-                                                in->getMemoryDesc(),
-                                                wgt->getMemoryDesc(),
-                                                out->getMemoryDesc());
-  fc_fwd::primitive_desc fwdPD = fc_fwd::primitive_desc(fwdDesc, engine_);
-  if (hasBias) {
-    fwd_.reset(new fc_fwd(fwdPD, *in, *wgt, *bias, *out));
+  fc_fwd::desc fwdDesc = bias != nullptr ? fc_fwd::desc(pk,
+                                                        in->getMemoryDesc(),
+                                                        wgt->getMemoryDesc(),
+                                                        bias->getMemoryDesc(),
+                                                        out->getMemoryDesc())
+                                         : fc_fwd::desc(pk,
+                                                        in->getMemoryDesc(),
+                                                        wgt->getMemoryDesc(),
+                                                        out->getMemoryDesc());
+  pd.reset(new fc_fwd::primitive_desc(fwdDesc, engine_));
+}
+
+void MKLDNNFcLayer::resetFwdPipeline(
+    std::vector<primitive>& pipeline,
+    std::shared_ptr<fc_fwd::primitive_desc>& pd,
+    MKLDNNMatrixPtr& in,
+    MKLDNNMatrixPtr& wgt,
+    MKLDNNMatrixPtr& bias,
+    MKLDNNMatrixPtr& out) {
+  pipeline.clear();
+
+  if (bias) {
+    fwd_.reset(new fc_fwd(*pd, *in, *wgt, *bias, *out));
   } else {
-    fwd_.reset(new fc_fwd(fwdPD, *in, *wgt, *out));
+    fwd_.reset(new fc_fwd(*pd, *in, *wgt, *out));
   }
-  printValueFormatFlow();
 
   pipeline.push_back(*fwd_);
 }
 
-void MKLDNNFcLayer::resetBwd(std::vector<mkldnn::primitive>& pipeline,
-                             MKLDNNMatrixPtr& in,
-                             MKLDNNMatrixPtr& wgt,
-                             MKLDNNMatrixPtr& bias,
-                             MKLDNNMatrixPtr& out) {
-  pipeline.clear();
-  if (!needResetBwd_) {
-    return;
-  }
-  needResetBwd_ = false;
-  bool hasBias = biases_ && biases_->getWGrad();
+void MKLDNNFcLayer::resetBwdBuffers(MKLDNNMatrixPtr& in,
+                                    MKLDNNMatrixPtr& wgt,
+                                    MKLDNNMatrixPtr& bias,
+                                    MKLDNNMatrixPtr& out) {
+  resetOutGrad(out);
+
+  resetWgtBiasGrad(wgt, bias);
 
-  /// backward weight
-  CHECK(inVal_) << "Should have input value";
-  const MatrixPtr& wgtGrad = weight_->getWGrad();
-  const MatrixPtr& biasGrad = hasBias ? biases_->getWGrad() : nullptr;
+  resetInGrad(in);
+}
 
+void MKLDNNFcLayer::resetOutGrad(MKLDNNMatrixPtr& out) {
   // TODO(TJ): merge outgrad
   int device = outputIsOnlyMKLDNN() ? MKLDNN_DEVICE : CPU_DEVICE;
   // for MKLDNN device:
@@ -178,66 +241,88 @@ void MKLDNNFcLayer::resetBwd(std::vector<mkldnn::primitive>& pipeline,
   // for CPU device:
   // fc do not need to convert from cpu device since output is always nc format
   // only need create from cpu device
-  const MatrixPtr& outGrad = getOutput(device).grad;
-  out = MKLDNNMatrix::create(outGrad, outVal_->getPrimitiveDesc());
-  wgt = MKLDNNMatrix::create(wgtGrad, wgtVal_->getPrimitiveDesc());
-  bias = hasBias ? MKLDNNMatrix::create(biasGrad, biasVal_->getPrimitiveDesc())
-                 : nullptr;
-
-  // create memory primitive desc
-  fc_fwd::desc fwdDesc = fc_fwd::desc(prop_kind::forward,
-                                      inVal_->getMemoryDesc(),
-                                      wgt->getMemoryDesc(),
-                                      out->getMemoryDesc());
-  fc_fwd::primitive_desc fwdPD = fc_fwd::primitive_desc(fwdDesc, engine_);
-  fc_bwdWgt::desc bwdWgtDesc = hasBias
-                                   ? fc_bwdWgt::desc(inVal_->getMemoryDesc(),
-                                                     wgt->getMemoryDesc(),
-                                                     bias->getMemoryDesc(),
-                                                     out->getMemoryDesc())
-                                   : fc_bwdWgt::desc(inVal_->getMemoryDesc(),
-                                                     wgt->getMemoryDesc(),
-                                                     out->getMemoryDesc());
-  fc_bwdWgt::primitive_desc bwdWgtPD =
-      fc_bwdWgt::primitive_desc(bwdWgtDesc, engine_, fwdPD);
-
-  if (hasBias) {
-    bwdWgt_.reset(new fc_bwdWgt(bwdWgtPD, *inVal_, *out, *wgt, *bias));
-  } else {
-    bwdWgt_.reset(new fc_bwdWgt(bwdWgtPD, *inVal_, *out, *wgt));
+  CHECK(outVal_);
+  out =
+      MKLDNNMatrix::create(getOutput(device).grad, outVal_->getPrimitiveDesc());
+}
+
+void MKLDNNFcLayer::resetWgtBiasGrad(MKLDNNMatrixPtr& wgt,
+                                     MKLDNNMatrixPtr& bias) {
+  CHECK(wgtVal_);
+  wgt = MKLDNNMatrix::create(weight_->getWGrad(), wgtVal_->getPrimitiveDesc());
+
+  bias = nullptr;
+  if (biasVal_ == nullptr) {
+    return;
   }
-  pipeline.push_back(*bwdWgt_);
+  bias =
+      MKLDNNMatrix::create(biases_->getWGrad(), biasVal_->getPrimitiveDesc());
+}
 
-  /// backward data
+void MKLDNNFcLayer::resetInGrad(MKLDNNMatrixPtr& in) {
+  in = nullptr;
   const MatrixPtr& inGrad = inputLayers_[0]->getOutput().grad;
   if (inGrad == nullptr) {
     return;
   }
-  if (getInput(0, MKLDNN_DEVICE).getAllCount() > 1) {
-    // TODO(TJ): use outputMaps_ ways to get the inGrad_ when merge outgrad done
-  } else {
-    in = MKLDNNMatrix::create(inGrad, inVal_->getPrimitiveDesc());
-  }
-
-  fc_bwdData::desc bwdDataDesc = fc_bwdData::desc(
-      inVal_->getMemoryDesc(), wgt->getMemoryDesc(), out->getMemoryDesc());
-  fc_bwdData::primitive_desc bwdDataPD =
-      fc_bwdData::primitive_desc(bwdDataDesc, engine_, fwdPD);
+  // TODO(TJ): use outputMaps_ ways to get the inGrad_ when merge outgrad done
+  CHECK(inVal_);
+  in = MKLDNNMatrix::create(inGrad, inVal_->getPrimitiveDesc());
+}
 
-  CHECK(wgtVal_) << "Should have weight memory";
-  bwdData_.reset(new fc_bwdData(bwdDataPD, *out, *wgtVal_, *in));
-  printGradFormatFlow();
-  pipeline.push_back(*bwdData_);
+void MKLDNNFcLayer::resetBwdWgtPD(
+    std::shared_ptr<fc_bwdWgt::primitive_desc>& pd,
+    MKLDNNMatrixPtr& wgt,
+    MKLDNNMatrixPtr& bias,
+    MKLDNNMatrixPtr& out) {
+  CHECK(inVal_);
+  fc_bwdWgt::desc bwdWgtDesc = bias ? fc_bwdWgt::desc(inVal_->getMemoryDesc(),
+                                                      wgt->getMemoryDesc(),
+                                                      bias->getMemoryDesc(),
+                                                      out->getMemoryDesc())
+                                    : fc_bwdWgt::desc(inVal_->getMemoryDesc(),
+                                                      wgt->getMemoryDesc(),
+                                                      out->getMemoryDesc());
+  pd.reset(new fc_bwdWgt::primitive_desc(bwdWgtDesc, engine_, *fwdPD_));
 }
 
-void MKLDNNFcLayer::updateInputData() {
-  inVal_->setData(getInputValue(0, CPU_DEVICE)->getData());
+void MKLDNNFcLayer::resetBwdDataPD(
+    std::shared_ptr<fc_bwdData::primitive_desc>& pd,
+    MKLDNNMatrixPtr& in,
+    MKLDNNMatrixPtr& out) {
+  pd = nullptr;
+  if (in == nullptr) {
+    return;
+  }
+  CHECK(wgtVal_);
+  fc_bwdData::desc bwdDataDesc = fc_bwdData::desc(
+      in->getMemoryDesc(), wgtVal_->getMemoryDesc(), out->getMemoryDesc());
+  pd.reset(new fc_bwdData::primitive_desc(bwdDataDesc, engine_, *fwdPD_));
 }
 
-void MKLDNNFcLayer::updateWeights(const UpdateCallback& callback) {
-  weight_->getParameterPtr()->incUpdate(callback);
-  if (biases_ && biases_->getWGrad()) {
-    biases_->getParameterPtr()->incUpdate(callback);
+void MKLDNNFcLayer::resetBwdPipeline(
+    std::vector<primitive>& pipeline,
+    std::shared_ptr<fc_bwdWgt::primitive_desc>& bwdWgtPD,
+    std::shared_ptr<fc_bwdData::primitive_desc>& bwdDataPD,
+    MKLDNNMatrixPtr& in,
+    MKLDNNMatrixPtr& wgt,
+    MKLDNNMatrixPtr& bias,
+    MKLDNNMatrixPtr& out) {
+  pipeline.clear();
+  CHECK(inVal_);
+  if (bias) {
+    bwdWgt_.reset(new fc_bwdWgt(*bwdWgtPD, *inVal_, *out, *wgt, *bias));
+  } else {
+    bwdWgt_.reset(new fc_bwdWgt(*bwdWgtPD, *inVal_, *out, *wgt));
+  }
+  pipeline.push_back(*bwdWgt_);
+
+  if (bwdDataPD == nullptr) {
+    return;
   }
+  CHECK(wgtVal_) << "Should have weight memory";
+  bwdData_.reset(new fc_bwdData(*bwdDataPD, *out, *wgtVal_, *in));
+  pipeline.push_back(*bwdData_);
 }
+
 }  // namespace paddle
diff --git a/paddle/gserver/layers/MKLDNNFcLayer.h b/paddle/gserver/layers/MKLDNNFcLayer.h
index 3119f863496df092da13c08bf733f13c42e53780..c76878aafab7e986d2bf478eaba02f2f0aced293 100644
--- a/paddle/gserver/layers/MKLDNNFcLayer.h
+++ b/paddle/gserver/layers/MKLDNNFcLayer.h
@@ -18,6 +18,9 @@ limitations under the License. */
 #include "mkldnn.hpp"
 
 namespace paddle {
+typedef mkldnn::inner_product_forward fc_fwd;
+typedef mkldnn::inner_product_backward_weights fc_bwdWgt;
+typedef mkldnn::inner_product_backward_data fc_bwdData;
 
 /**
  * @brief A subclass of MKLDNNLayer fc layer.
@@ -32,6 +35,9 @@ protected:
   // if has already init the weight
   bool hasInitedWgt_;
 
+  // save forward primitive_desc, which can be used backward
+  std::shared_ptr<fc_fwd::primitive_desc> fwdPD_;
+
   // fc weight and bias
   std::unique_ptr<Weight> weight_;
   std::unique_ptr<Weight> biases_;
@@ -67,6 +73,59 @@ public:
   void convertWeightsFromPaddle() override;
 
   void convertWeightsToPaddle() override;
+
+protected:
+  /**
+   * Forward functions: reset buffers(input, output, weight and bias),
+   *                    reset primitive descriptor,
+   *                    reset pipeline.
+   */
+  void resetFwdBuffers(MKLDNNMatrixPtr& in,
+                       MKLDNNMatrixPtr& wgt,
+                       MKLDNNMatrixPtr& bias,
+                       MKLDNNMatrixPtr& out);
+  void resetInValue(MKLDNNMatrixPtr& in);
+  void resetWgtBiasValue(MKLDNNMatrixPtr& wgt, MKLDNNMatrixPtr& bias);
+  void resetOutValue(MKLDNNMatrixPtr& out);
+  void resetFwdPD(std::shared_ptr<fc_fwd::primitive_desc>& pd,
+                  MKLDNNMatrixPtr in,
+                  MKLDNNMatrixPtr wgt,
+                  MKLDNNMatrixPtr bias,
+                  MKLDNNMatrixPtr out);
+  void resetFwdPipeline(std::vector<mkldnn::primitive>& pipeline,
+                        std::shared_ptr<fc_fwd::primitive_desc>& pd,
+                        MKLDNNMatrixPtr& in,
+                        MKLDNNMatrixPtr& wgt,
+                        MKLDNNMatrixPtr& bias,
+                        MKLDNNMatrixPtr& out);
+
+  /**
+   * Backward functions: reset buffers(input, output, weight and bias),
+   *                     reset primitive descriptor for backward weight,
+   *                     reset primitive descriptor for backward data,
+   *                     reset pipeline.
+   */
+  void resetBwdBuffers(MKLDNNMatrixPtr& in,
+                       MKLDNNMatrixPtr& wgt,
+                       MKLDNNMatrixPtr& bias,
+                       MKLDNNMatrixPtr& out);
+  void resetOutGrad(MKLDNNMatrixPtr& out);
+  void resetWgtBiasGrad(MKLDNNMatrixPtr& wgt, MKLDNNMatrixPtr& bias);
+  void resetInGrad(MKLDNNMatrixPtr& in);
+  void resetBwdWgtPD(std::shared_ptr<fc_bwdWgt::primitive_desc>& pd,
+                     MKLDNNMatrixPtr& wgt,
+                     MKLDNNMatrixPtr& bias,
+                     MKLDNNMatrixPtr& out);
+  void resetBwdDataPD(std::shared_ptr<fc_bwdData::primitive_desc>& pd,
+                      MKLDNNMatrixPtr& in,
+                      MKLDNNMatrixPtr& out);
+  void resetBwdPipeline(std::vector<mkldnn::primitive>& pipeline,
+                        std::shared_ptr<fc_bwdWgt::primitive_desc>& bwdWgtPD,
+                        std::shared_ptr<fc_bwdData::primitive_desc>& bwdDataPD,
+                        MKLDNNMatrixPtr& in,
+                        MKLDNNMatrixPtr& wgt,
+                        MKLDNNMatrixPtr& bias,
+                        MKLDNNMatrixPtr& out);
 };
 
 }  // namespace paddle
diff --git a/paddle/gserver/tests/test_MKLDNN.cpp b/paddle/gserver/tests/test_MKLDNN.cpp
index e1d2270df24331914f3a51acc90a518084b3ce4e..e70802881e3f22160a87b7a4babda07ffbcf9d6f 100644
--- a/paddle/gserver/tests/test_MKLDNN.cpp
+++ b/paddle/gserver/tests/test_MKLDNN.cpp
@@ -17,6 +17,7 @@ limitations under the License. */
 #include <vector>
 #include "MKLDNNTester.h"
 #include "ModelConfig.pb.h"
+#include "paddle/math/MathUtils.h"
 
 using namespace paddle;  // NOLINT
 
@@ -63,6 +64,83 @@ TEST(MKLDNNLayer, FcLayer) {
   testFcLayer({/*bs*/ 15, /*ic*/ 3, /*oc*/ 6, /*ih*/ 16, /*iw*/ 16});
 }
 
+struct testConvDesc {
+  int bs, gp;
+  int ic, ih, iw;
+  int oc, oh, ow;
+  int fh, fw;
+  int ph, pw;
+  int sh, sw;
+  int dh, dw;
+};
+
+void testConvLayer(const testConvDesc& pm) {
+  const std::string compareTypes[] = {"mkldnn_conv", "exconv"};
+  TestConfig cfg;
+  cfg.layerConfig.set_type(compareTypes[0]);
+  cfg.layerConfig.set_num_filters(pm.oc);
+  cfg.layerConfig.set_size(pm.oc * pm.oh * pm.ow);
+  // cfg.layerConfig.set_partial_sum(1); // TODO: check it
+  cfg.layerConfig.set_shared_biases(true);
+  cfg.inputDefs.push_back(
+      {INPUT_DATA,
+       "layer_0",
+       /* size of input layer= */ size_t(pm.ic * pm.ih * pm.iw),
+       /* size of weight= */ size_t(pm.oc * pm.ic * pm.fh * pm.fw / pm.gp)});
+  LayerInputConfig* input = cfg.layerConfig.add_inputs();
+  ConvConfig* conv = input->mutable_conv_conf();
+  conv->set_groups(pm.gp);
+  conv->set_img_size(pm.iw);
+  conv->set_img_size_y(pm.ih);
+  conv->set_output_x(pm.ow);
+  conv->set_output_y(pm.oh);
+  conv->set_filter_size(pm.fw);
+  conv->set_filter_size_y(pm.fh);
+  conv->set_channels(pm.ic);
+  conv->set_padding(pm.pw);
+  conv->set_padding_y(pm.ph);
+  conv->set_stride(pm.sw);
+  conv->set_stride_y(pm.sh);
+  conv->set_dilation(pm.dw);
+  conv->set_dilation_y(pm.dh);
+  conv->set_caffe_mode(true);
+  conv->set_filter_channels(conv->channels() / conv->groups());
+  CHECK_EQ(conv->filter_channels() * pm.gp, conv->channels())
+      << "it is indivisible";
+
+  int fh = (pm.fh - 1) * pm.dh + 1;
+  int fw = (pm.fw - 1) * pm.dw + 1;
+  int ow = outputSize(pm.iw, fw, pm.pw, pm.sw, true);
+  int oh = outputSize(pm.ih, fh, pm.ph, pm.sh, true);
+  CHECK_EQ(ow, pm.ow) << "output size check failed";
+  CHECK_EQ(oh, pm.oh) << "output size check failed";
+
+  MKLDNNTester tester;
+  for (auto biasSize : {pm.oc, 0}) {
+    cfg.biasSize = biasSize;
+    TestConfig ref = cfg;
+    ref.layerConfig.set_type(compareTypes[1]);
+    for (auto bs : {pm.bs, 1}) {
+      tester.run(cfg, ref, bs, pm.ih, pm.iw);
+    }
+  }
+}
+
+TEST(MKLDNNLayer, ConvLayer) {
+  /* bs, gp, ic, ih, iw, oc, oh, ow, fh, fw, ph, pw, sh, sw, dh, dw */
+  testConvLayer({2, 1, 3, 32, 32, 16, 32, 32, 3, 3, 1, 1, 1, 1, 1, 1});
+  testConvLayer({2, 1, 8, 16, 16, 8, 16, 16, 3, 3, 1, 1, 1, 1, 1, 1});
+  testConvLayer({3, 1, 16, 32, 32, 3, 32, 32, 3, 3, 1, 1, 1, 1, 1, 1});
+  testConvLayer({8, 1, 16, 18, 18, 32, 18, 18, 3, 3, 1, 1, 1, 1, 1, 1});
+  testConvLayer({16, 1, 1, 42, 31, 32, 23, 11, 4, 5, 3, 2, 2, 3, 1, 1});
+  testConvLayer({2, 1, 8, 16, 16, 8, 8, 8, 3, 3, 1, 1, 2, 2, 1, 1});
+  testConvLayer({3, 1, 8, 13, 13, 8, 7, 7, 3, 3, 1, 1, 2, 2, 1, 1});
+  // with groups
+  testConvLayer({2, 2, 4, 5, 5, 8, 5, 5, 3, 3, 1, 1, 1, 1, 1, 1});
+  testConvLayer({2, 3, 3, 5, 5, 3, 5, 5, 3, 3, 1, 1, 1, 1, 1, 1});
+  testConvLayer({4, 4, 16, 3, 3, 16, 3, 3, 3, 3, 1, 1, 1, 1, 1, 1});
+}
+
 // TODO(TJ): add branch test
 
 int main(int argc, char** argv) {
diff --git a/paddle/math/MKLDNNMatrix.cpp b/paddle/math/MKLDNNMatrix.cpp
index c4063e5069854242d9f93886b66580385557ca73..0778bb63b7b3bca9b3d2647ca43dad72d783950a 100644
--- a/paddle/math/MKLDNNMatrix.cpp
+++ b/paddle/math/MKLDNNMatrix.cpp
@@ -49,6 +49,27 @@ MKLDNNMatrixPtr MKLDNNMatrix::create(MatrixPtr m,
   return create(m, memory::primitive_desc(memory::desc(dims, dtype, fmt), eg));
 }
 
+std::shared_ptr<reorder> MKLDNNMatrix::createReorder(const MKLDNNMatrixPtr& src,
+                                                     const MKLDNNMatrixPtr& dst,
+                                                     bool checkData) {
+  if (src == dst || src->getPrimitiveDesc() == dst->getPrimitiveDesc()) {
+    return nullptr;
+  }
+
+  if (checkData && (src->getData() == dst->getData())) {
+    LOG(FATAL) << "can not create reorder with inplace data";
+    return nullptr;
+  }
+
+  memory::dims srcDims = src->getDims();
+  memory::dims dstDims = dst->getDims();
+  CHECK_EQ(srcDims.size(), dstDims.size());
+  for (size_t i = 0; i < srcDims.size(); ++i) {
+    CHECK_EQ(srcDims[i], dstDims[i]);
+  }
+  return std::make_shared<reorder>(*src, *dst);
+}
+
 void MKLDNNMatrix::reorderDataFrom(const MKLDNNMatrixPtr& m,
                                    memory::format srcFmt,
                                    memory::dims targetDim) {
diff --git a/paddle/math/MKLDNNMatrix.h b/paddle/math/MKLDNNMatrix.h
index eef3b429e6fa0087aeac3f5aed9dff983b06e826..c843115eb9a5be50d6ff873f1510844228c9d89f 100644
--- a/paddle/math/MKLDNNMatrix.h
+++ b/paddle/math/MKLDNNMatrix.h
@@ -52,6 +52,32 @@ public:
       mkldnn::engine& eg,
       mkldnn::memory::data_type dtype = mkldnn::memory::data_type::f32);
 
+  /**
+   * Create Memory descriptor.
+   * default with any format and f32 dtype
+   */
+  static mkldnn::memory::desc createMemoryDesc(
+      const mkldnn::memory::dims& dims,
+      const mkldnn::memory::format& fmt = mkldnn::memory::format::any,
+      const mkldnn::memory::data_type& dtype = mkldnn::memory::data_type::f32) {
+    return mkldnn::memory::desc(dims, dtype, fmt);
+  }
+
+  /**
+   * Create reorder primitive.
+   * Create a mkldnn::reorder handle for converting src MKLDNNMatrix to dst.
+   * checkData: whether to check the data handle of src and dst.
+   *            if true, it will check the data and do not allow them equal;
+   *            otherwise, it will not check them, then the reorder created
+   *            may have inplace buffer.
+   *            Do not set false, if you can not guarantee the inplace logical
+   *            would work with your reorder.
+   */
+  static std::shared_ptr<mkldnn::reorder> createReorder(
+      const MKLDNNMatrixPtr& src,
+      const MKLDNNMatrixPtr& dst,
+      bool checkData = true);
+
 public:
   /**
    * Reorder this MKLDNNMatrix from other format.
diff --git a/paddle/memory/memcpy.cc b/paddle/memory/memcpy.cc
index a19a3e3675e3e2e7cc0c3594f21191f932d6379f..19ec9ba9b26f5919796181a19a048b7edb508bdd 100644
--- a/paddle/memory/memcpy.cc
+++ b/paddle/memory/memcpy.cc
@@ -62,6 +62,24 @@ void Copy<platform::GPUPlace, platform::GPUPlace>(platform::GPUPlace dst_place,
   }
 }
 
+template <>
+void Copy<platform::CPUPlace, platform::GPUPlace>(platform::CPUPlace dst_place,
+                                                  void* dst,
+                                                  platform::GPUPlace src_place,
+                                                  const void* src, size_t num) {
+  platform::SetDeviceId(src_place.device);
+  platform::GpuMemcpySync(dst, src, num, cudaMemcpyDeviceToHost);
+}
+
+template <>
+void Copy<platform::GPUPlace, platform::CPUPlace>(platform::GPUPlace dst_place,
+                                                  void* dst,
+                                                  platform::CPUPlace src_place,
+                                                  const void* src, size_t num) {
+  platform::SetDeviceId(dst_place.device);
+  platform::GpuMemcpySync(dst, src, num, cudaMemcpyHostToDevice);
+}
+
 #endif  // PADDLE_ONLY_CPU
 
 }  // namespace memory
diff --git a/paddle/operators/CMakeLists.txt b/paddle/operators/CMakeLists.txt
index f9ea25ab045a02be5ab9ed81ef9c679126d3a188..e3e934bcccd1a5f34d88a2f33f3708a46ddabe05 100644
--- a/paddle/operators/CMakeLists.txt
+++ b/paddle/operators/CMakeLists.txt
@@ -1,5 +1,7 @@
 file(GLOB GENERAL_OPS RELATIVE "${CMAKE_CURRENT_SOURCE_DIR}" "*_op.cc")
 string(REPLACE ".cc" "" GENERAL_OPS "${GENERAL_OPS}")
+set(pybind_file ${PADDLE_SOURCE_DIR}/paddle/pybind/pybind.h)
+file(WRITE ${pybind_file} "// Generated by the paddle/operator/CMakeLists.txt.  DO NOT EDIT!\n\n")
 function(op_library TARGET)
     # op_library is a function to create op library. The interface is same as
     # cc_library. But it handle split GPU/CPU code and link some common library
@@ -7,10 +9,11 @@ function(op_library TARGET)
     set(OP_LIBRARY ${TARGET} ${OP_LIBRARY} PARENT_SCOPE)
     set(cc_srcs)
     set(cu_srcs)
-    set(op_common_deps operator op_registry)
+    set(op_common_deps operator op_registry math_function)
     set(options "")
     set(oneValueArgs "")
     set(multiValueArgs SRCS DEPS)
+    set(pybind_flag 0)
     cmake_parse_arguments(op_library "${options}" "${oneValueArgs}"
             "${multiValueArgs}" ${ARGN})
 
@@ -46,22 +49,42 @@ function(op_library TARGET)
         cc_library(${TARGET} SRCS ${cc_srcs} DEPS ${op_library_DEPS}
                 ${op_common_deps})
     endif()
+
+    # net_op doesn't need pybind
+    if ("${TARGET}" STREQUAL "net_op")
+        set(pybind_flag 1)
+    endif()
+
+    # pybind USE_NO_KERNEL_OP
+    file(READ ${TARGET}.cc TARGET_CONTENT)
+    string(REGEX MATCH "OperatorWithKernel" regex_result "${TARGET_CONTENT}")
+    string(REPLACE "_op" "" TARGET "${TARGET}")
+    if (${pybind_flag} EQUAL 0 AND regex_result STREQUAL "")
+        file(APPEND ${pybind_file} "USE_NO_KERNEL_OP(${TARGET});\n")
+        set(pybind_flag 1)
+    endif()
+
+    # pybind USE_CPU_ONLY_OP
+    list(LENGTH cu_srcs cu_srcs_len)
+    if (${pybind_flag} EQUAL 0 AND ${cu_srcs_len} EQUAL 0)
+        file(APPEND ${pybind_file} "USE_CPU_ONLY_OP(${TARGET});\n")
+        set(pybind_flag 1)
+    endif()
+
+    # pybind USE_OP
+    if (${pybind_flag} EQUAL 0)
+        file(APPEND ${pybind_file} "USE_OP(${TARGET});\n")
+    endif()
 endfunction()
 
 add_subdirectory(math)
 
 set(DEPS_OPS
-    identity_op
-    minus_op
-    mul_op
     recurrent_op
-    scale_op)
-op_library(identity_op DEPS scale_op)
-op_library(minus_op DEPS scale_op)
-op_library(mul_op DEPS math_function)
+    cond_op)
 op_library(recurrent_op SRCS recurrent_op.cc rnn/recurrent_op_utils.cc
-  DEPS framework_proto tensor operator net_op)
-op_library(scale_op DEPS net_op)
+  DEPS framework_proto tensor net_op)
+op_library(cond_op SRCS cond_op.cc DEPS framework_proto tensor operator net_op)
 
 list(REMOVE_ITEM GENERAL_OPS ${DEPS_OPS})
 foreach(src ${GENERAL_OPS})
diff --git a/paddle/operators/accuracy_op.cc b/paddle/operators/accuracy_op.cc
new file mode 100644
index 0000000000000000000000000000000000000000..0c813748b2989a8f0c00a359345747242dd21dd8
--- /dev/null
+++ b/paddle/operators/accuracy_op.cc
@@ -0,0 +1,71 @@
+/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+    http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License. */
+
+#include "paddle/operators/accuracy_op.h"
+
+namespace paddle {
+namespace operators {
+
+class AccuracyOp : public framework::OperatorWithKernel {
+ public:
+  using framework::OperatorWithKernel::OperatorWithKernel;
+
+ protected:
+  void InferShape(const framework::InferShapeContext &ctx) const override {
+    PADDLE_ENFORCE_NOT_NULL(
+        ctx.InputVar("Inference"),
+        "Input(Inference) of AccuracyOp should not be null.");
+    PADDLE_ENFORCE_NOT_NULL(ctx.InputVar("Label"),
+                            "Input(Label) of AccuracyOp should not be null.");
+    PADDLE_ENFORCE_NOT_NULL(
+        ctx.OutputVar("Accuracy"),
+        "Output(Accuracy) of AccuracyOp should not be null.");
+
+    auto *inference = ctx.Input<framework::Tensor>("Inference");
+    auto *label = ctx.Input<framework::Tensor>("Label");
+
+    PADDLE_ENFORCE_EQ(label->dims().size(), 1, "label must be a vector");
+    PADDLE_ENFORCE_EQ(inference->dims()[0], label->dims()[0],
+                      "inference size must be the same as label size");
+
+    ctx.Output<framework::LoDTensor>("Accuracy")->Resize({1});
+  }
+};
+
+class AccuracyOpMaker : public framework::OpProtoAndCheckerMaker {
+ public:
+  AccuracyOpMaker(framework::OpProto *proto,
+                  framework::OpAttrChecker *op_checker)
+      : OpProtoAndCheckerMaker(proto, op_checker) {
+    // TODO(typhoonzero): support both inference value and indices.
+    AddInput("Inference", "topk(indices) the network output");
+    AddInput("Label", "Label of the training data");
+    // TODO(typhoonzero): AddInput("Weight", ...
+    AddOutput("Accuracy", "The accuracy of current batch");
+
+    AddComment(
+        R"DOC(Accuracy. It will print accuracy rate for classification.
+The accuracy is:
+..  math::
+accuracy = \\frac{NumOfCorrectPredicts}{NumOfAllSamples})DOC");
+  }
+};
+
+}  // namespace operators
+}  // namespace paddle
+
+namespace ops = paddle::operators;
+REGISTER_OP_WITHOUT_GRADIENT(accuracy, ops::AccuracyOp, ops::AccuracyOpMaker);
+REGISTER_OP_CPU_KERNEL(accuracy,
+                       ops::AccuracyKernel<paddle::platform::CPUPlace, float>);
diff --git a/paddle/operators/accuracy_op.cu b/paddle/operators/accuracy_op.cu
new file mode 100644
index 0000000000000000000000000000000000000000..4e6d1ef9654012ce6355cbd7561c4fdc1785c11a
--- /dev/null
+++ b/paddle/operators/accuracy_op.cu
@@ -0,0 +1,69 @@
+/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+    http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License. */
+
+#include "paddle/operators/accuracy_op.h"
+
+namespace paddle {
+namespace operators {
+
+__global__ void AccuracySingleKernel(const int N, const int D, const int top_k,
+                                     const int* Xdata, const int* labelData,
+                                     float* accuracy) {
+  int correct = 0;
+  for (int row = 0; row < N; row++) {
+    const int label = labelData[row];
+    for (int col = 0; col < D; col++) {
+      const int pred = Xdata[row * D + col];
+      if (pred == label) {
+        ++correct;
+        break;
+      }
+    }
+  }
+  *accuracy = static_cast<float>(correct) / static_cast<float>(N);
+}
+
+template <typename T>
+class AccuracyOpCUDAKernel : public framework::OpKernel {
+ public:
+  void Compute(const framework::ExecutionContext& ctx) const override {
+    PADDLE_ENFORCE(platform::is_gpu_place(ctx.GetPlace()),
+                   "It must use GPUPlace.");
+    auto* inference = ctx.Input<Tensor>("Inference");
+    auto* label = ctx.Input<Tensor>("Label");
+    auto* accuracy = ctx.Output<Tensor>("Accuracy");
+    // FIXME(typhoonzero): only support indices currently
+    // if add support for output values, how to detect the data type?
+    const int* inference_data = inference->data<int>();
+    const int* label_data = label->data<int>();
+    float* accuracy_data = accuracy->mutable_data<float>(ctx.GetPlace());
+
+    size_t num_samples = inference->dims()[0];
+    size_t infer_width = inference->dims()[1];
+    cudaMemset((void**)&accuracy_data, 0, sizeof(float));
+
+    if (num_samples == 0) {
+      return;
+    }
+
+    AccuracySingleKernel<<<1, 1>>>(num_samples, infer_width, 1, inference_data,
+                                   label_data, accuracy_data);
+  }
+};
+
+}  // namespace operators
+}  // namespace paddle
+
+REGISTER_OP_GPU_KERNEL(accuracy,
+                       paddle::operators::AccuracyOpCUDAKernel<float>);
diff --git a/paddle/operators/accuracy_op.h b/paddle/operators/accuracy_op.h
new file mode 100644
index 0000000000000000000000000000000000000000..fe704efe1c979f4fc6a5a37184e51b416f5e517f
--- /dev/null
+++ b/paddle/operators/accuracy_op.h
@@ -0,0 +1,77 @@
+/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+    http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License. */
+
+#pragma once
+#include <algorithm>
+#include "paddle/framework/eigen.h"
+#include "paddle/framework/op_registry.h"
+
+namespace paddle {
+namespace operators {
+
+using Tensor = framework::Tensor;
+
+template <typename T, int MajorType = Eigen::RowMajor,
+          typename IndexType = Eigen::DenseIndex>
+using EigenMatrix = framework::EigenMatrix<T, MajorType, IndexType>;
+
+template <typename T, int MajorType = Eigen::RowMajor,
+          typename IndexType = Eigen::DenseIndex>
+using EigenVector = framework::EigenVector<T, MajorType, IndexType>;
+
+template <typename T, int MajorType = Eigen::RowMajor,
+          typename IndexType = Eigen::DenseIndex>
+using EigenScalar = framework::EigenScalar<T, MajorType, IndexType>;
+
+template <typename Place, typename T>
+class AccuracyKernel : public framework::OpKernel {
+ public:
+  void Compute(const framework::ExecutionContext& ctx) const override {
+    auto* inference = ctx.Input<Tensor>("Inference");
+    auto* label = ctx.Input<Tensor>("Label");
+    auto* accuracy = ctx.Output<Tensor>("Accuracy");
+
+    float* accuracy_data = accuracy->mutable_data<float>(ctx.GetPlace());
+
+    const T* inference_data = inference->data<T>();
+    const T* label_data = label->data<T>();
+
+    size_t num_samples = inference->dims()[0];
+    size_t class_dim = inference->dims()[1];
+    *accuracy_data = 0.0f;
+
+    if (num_samples == 0) {
+      return;
+    }
+
+    int num_correct = 0;
+    // assume inference is already the topk of the output
+    for (size_t i = 0; i < num_samples; ++i) {
+      PADDLE_ENFORCE_GE(label_data[i], 0, "label must >= 0");
+      for (size_t j = 0; j < class_dim; ++j) {
+        if (inference_data[i * class_dim + j] == label_data[i]) {
+          ++num_correct;
+          break;
+        }
+      }
+    }
+
+    // FIXME(typhoonzero): we don't accumulate the accuracy for now.
+    *accuracy_data =
+        static_cast<float>(num_correct) / static_cast<float>(num_samples);
+  }
+};
+
+}  // namespace operators
+}  // namespace paddle
diff --git a/paddle/operators/add_op.cc b/paddle/operators/add_op.cc
index 8dbd47cf0dfbc265032a9966343eed5c7bd8692e..e83c1efeaf897889d18a37a6bd2ca2f8f012db25 100644
--- a/paddle/operators/add_op.cc
+++ b/paddle/operators/add_op.cc
@@ -23,10 +23,18 @@ class AddOp : public framework::OperatorWithKernel {
 
  protected:
   void InferShape(const framework::InferShapeContext &ctx) const override {
+    PADDLE_ENFORCE_NOT_NULL(ctx.InputVar("X"),
+                            "Input(X) of AddOp should not be null.");
+    PADDLE_ENFORCE_NOT_NULL(ctx.InputVar("Y"),
+                            "Input(Y) of AddOp should not be null.");
+    PADDLE_ENFORCE_NOT_NULL(ctx.OutputVar("Out"),
+                            "Output(Out) of AddOp should not be null.");
+
     PADDLE_ENFORCE_EQ(ctx.Input<Tensor>("X")->dims(),
                       ctx.Input<Tensor>("Y")->dims(),
                       "Two input of Add Op's dimension must be same.");
-    ctx.Output<Tensor>("Out")->Resize(ctx.Input<Tensor>("X")->dims());
+    ctx.Output<framework::LoDTensor>("Out")->Resize(
+        ctx.Input<Tensor>("X")->dims());
   }
 };
 
diff --git a/paddle/operators/concat_op.cc b/paddle/operators/concat_op.cc
index 0ebefbab26ec8fdf316f852fbb7f6d9f3bbc48eb..223bb0ffe6e75ce71919eb5f4cca06bedbb00764 100644
--- a/paddle/operators/concat_op.cc
+++ b/paddle/operators/concat_op.cc
@@ -25,8 +25,11 @@ class ConcatOp : public framework::OperatorWithKernel {
 
  protected:
   void InferShape(const framework::InferShapeContext &ctx) const override {
+    PADDLE_ENFORCE_NOT_NULL(ctx.OutputVar("Out"),
+                            "Output(Out) of ConcatOp should not be null.");
+
     auto ins = ctx.MultiInput<framework::Tensor>("X");
-    auto *out = ctx.Output<framework::Tensor>("Out");
+    auto *out = ctx.Output<framework::LoDTensor>("Out");
     size_t axis = static_cast<size_t>(ctx.Attr<int>("axis"));
     size_t n = ins.size();
 
diff --git a/paddle/operators/concat_op.cu b/paddle/operators/concat_op.cu
deleted file mode 100644
index 38fee7473dbb2ba97fe95b6632db7a1749cf3bbe..0000000000000000000000000000000000000000
--- a/paddle/operators/concat_op.cu
+++ /dev/null
@@ -1,19 +0,0 @@
-/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve.
-
-Licensed under the Apache License, Version 2.0 (the "License");
-you may not use this file except in compliance with the License.
-You may obtain a copy of the License at
-
-http://www.apache.org/licenses/LICENSE-2.0
-
-Unless required by applicable law or agreed to in writing, software
-distributed under the License is distributed on an "AS IS" BASIS,
-WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-See the License for the specific language governing permissions and
-limitations under the License. */
-
-#define EIGEN_USE_GPU
-#include "paddle/operators/concat_op.h"
-
-namespace ops = paddle::operators;
-// TODO(Yancey1989) Add GPU kernel
diff --git a/paddle/operators/cond_op.cc b/paddle/operators/cond_op.cc
new file mode 100644
index 0000000000000000000000000000000000000000..8262a7a5c8c13c86c5f6c123a14fa89696358c57
--- /dev/null
+++ b/paddle/operators/cond_op.cc
@@ -0,0 +1,229 @@
+/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+    http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License. */
+
+#include "paddle/operators/cond_op.h"
+
+#include <cstring>
+#include <sstream>
+
+#include "paddle/framework/op_registry.h"
+#include "paddle/operators/gather.h"
+#include "paddle/operators/net_op.h"
+#include "paddle/operators/scatter.h"
+
+namespace paddle {
+namespace operators {
+
+using Scope = framework::Scope;
+using Variable = framework::Variable;
+using Tensor = framework::Tensor;
+using LoDTensor = framework::LoDTensor;
+using DDim = framework::DDim;
+
+void CondOp::CreateScope(const Scope& scope) const {
+  auto sub_scopes_var = scope.FindVar("SubScopes");
+  PADDLE_ENFORCE_NOT_NULL(sub_scopes_var,
+                          "Output(SubScopes) of CondOp should not be null.");
+  auto sub_scopes = sub_scopes_var->GetMutable<std::vector<Scope*>>();
+  auto& sub_scope = scope.NewScope();
+  sub_scopes->push_back(&sub_scope);
+}
+
+void CondOp::CreateIndexTensor(const Scope& scope) const {
+  auto index_tensors_var = scope.FindVar("IndexTensors");
+  PADDLE_ENFORCE_NOT_NULL(index_tensors_var,
+                          "Output(IndexTensors) of CondOp should not be null.");
+  auto& index_tensors =
+      *index_tensors_var->GetMutable<std::vector<LoDTensor>>();
+  index_tensors.push_back(LoDTensor());
+}
+
+void CondOp::InferShape(const Scope& scope) const {
+  auto sub_scopes_var = scope.FindVar("SubScopes");
+  PADDLE_ENFORCE_NOT_NULL(sub_scopes_var,
+                          "Output(SubScopes) of CondOp should not be null.");
+  auto& sub_scopes = *sub_scopes_var->GetMutable<std::vector<Scope*>>();
+
+  for (int i = 0; i < 2; ++i) {
+    // Create two sub scopes for true and false branches
+    // sub_scopes[0] for the true branch and sub_scopes[1] for the false
+    // branch
+    CreateScope(scope);
+
+    // Create two tensors for true and false indices
+    // index_tensors[0] for the true branch and index_tensors[1] for the false
+    // branch
+    CreateIndexTensor(scope);
+
+    PADDLE_ENFORCE(!Inputs("Xs").empty(),
+                   "Inputs(Xs) of CondOp can't be empty.");
+    for (auto& input : Inputs("Xs")) {
+      // Create a new tensor in sub-scope for input-type tensor
+      Variable* v = sub_scopes[i]->NewVar(input);
+      LoDTensor* sub_input = v->GetMutable<LoDTensor>();
+      sub_input->Resize(scope.FindVar(input)->GetMutable<LoDTensor>()->dims());
+    }
+
+    for (auto& output : (*sub_net_op_[i]).Outputs()) {
+      for (auto& var_name : output.second) {
+        sub_scopes[i]->NewVar(var_name);
+      }
+    }
+
+    // each net calls InferShape
+    sub_net_op_[i]->InferShape(*sub_scopes[i]);
+  }
+
+  for (auto& output : Outputs("Outs")) {
+    LoDTensor* tensor_t_out =
+        sub_scopes[0]->FindVar(output)->GetMutable<LoDTensor>();
+    PADDLE_ENFORCE_NOT_NULL(tensor_t_out, "True output should not be NULL");
+    LoDTensor* tensor_f_out =
+        sub_scopes[1]->FindVar(output)->GetMutable<LoDTensor>();
+    PADDLE_ENFORCE_NOT_NULL(tensor_f_out, "False output should not be NULL");
+
+    auto* tensor_out_var = scope.FindVar(output);
+    PADDLE_ENFORCE_NOT_NULL(tensor_out_var, "Output not found");
+    LoDTensor* tensor_out = tensor_out_var->GetMutable<LoDTensor>();
+    PADDLE_ENFORCE_NOT_NULL(tensor_t_out,
+                            "True output tensor should not be NULL");
+
+    // check output size should be same
+    PADDLE_ENFORCE_EQ(tensor_t_out->dims(), tensor_f_out->dims(),
+                      "Outputs not of the same shape");
+    tensor_out->Resize(tensor_t_out->dims());
+    // tensor_out->mutable_data<float>(tensor_out->dims(),
+    // platform::CPUPlace());
+    tensor_out->mutable_data<float>(platform::CPUPlace());
+  }
+}
+
+void CondOp::Run(const Scope& scope,
+                 const platform::DeviceContext& dev_ctx) const {
+  auto* sub_scopes_var = scope.FindVar("SubScopes");
+  PADDLE_ENFORCE_NOT_NULL(sub_scopes_var,
+                          "Output(SubScopes) of CondOp should not be null.");
+  auto sub_scopes = sub_scopes_var->Get<std::vector<Scope*>>();
+  auto* index_tensors_var = scope.FindVar("IndexTensors");
+  PADDLE_ENFORCE_NOT_NULL(index_tensors_var,
+                          "Output(IndexTensors) of CondOp should not be null.");
+  auto index_tensors = index_tensors_var->Get<std::vector<LoDTensor>>();
+
+  std::string cond_name = Input("Cond");
+  Variable* cond_var = scope.FindVar(cond_name);
+  PADDLE_ENFORCE_NOT_NULL(cond_var,
+                          "Input(Cond) of CondOp should not be null.");
+  const LoDTensor* cond = cond_var->GetMutable<LoDTensor>();
+
+  // Step 1: get the true/false index at runtime
+  // index_[0]: vector<int>, contains all index for cond[i] == true
+  // index_[1]: vector<int>, contains all index for cond[i] == false
+  for (int i = 0; i < 2; ++i) index_[i].clear();
+
+  const int* cond_data = cond->data<int>();
+  for (int i = 0; i < cond->dims()[0]; ++i) {
+    if (cond_data[i])
+      index_[0].push_back(i);
+    else
+      index_[1].push_back(i);
+  }
+
+  // put index_[0] and index_[1] into two tensors:
+  // index_tensor_[0] and index_tensor_[1]
+  DDim dim = paddle::framework::make_ddim({0});
+  for (int i = 0; i < 2; ++i) {
+    dim[0] = index_[i].size();
+    int* tmp_ptr =
+        index_tensors[i].mutable_data<int>(dim, platform::CPUPlace());
+    index_tensors[i].Resize(dim);
+    memcpy(tmp_ptr, index_[i].data(), dim[0] * sizeof(int));
+  }
+
+  // Step 2: collect data by calling gather
+  for (int i = 0; i < 2; ++i) {
+    // i= 0/i for True and False branches respectively
+    for (auto& input : Inputs("Xs")) {
+      // find Tensor
+      Variable* v = scope.FindVar(input);
+      PADDLE_ENFORCE_NOT_NULL(v);
+      LoDTensor* tensor_parent = v->GetMutable<LoDTensor>();
+
+      v = sub_scopes[i]->FindVar(input);
+      PADDLE_ENFORCE_NOT_NULL(v);
+      LoDTensor* tensor_child = v->GetMutable<LoDTensor>();
+
+      // Resize child
+      DDim dim = tensor_child->dims();
+      dim[0] = index_[i].size();
+      tensor_child->Resize(dim);
+      tensor_child->mutable_data<float>(dim, platform::CPUPlace());
+
+      Gather<float>(dev_ctx.GetPlace(), tensor_parent, &index_tensors[i],
+                    tensor_child);
+    }
+  }
+
+  // Step 3: run
+  for (int i = 0; i < 2; ++i) {
+    sub_net_op_[i]->Run(*sub_scopes[i], dev_ctx);
+  }
+
+  // Step 4: merge output results
+  PADDLE_ENFORCE(!Outputs("Outs").empty(),
+                 "Outputs(Outs) of CondOp can't be empty.");
+  for (int i = 0; i < 2; ++i) {
+    // i= 0/i for True and False branches respectively
+    for (auto& output : Outputs("Outs")) {
+      // find Tensor
+      Variable* v = scope.FindVar(output);
+      PADDLE_ENFORCE_NOT_NULL(v);
+      LoDTensor* tensor_parent = v->GetMutable<LoDTensor>();
+
+      v = sub_scopes[i]->FindVar(output);
+      PADDLE_ENFORCE_NOT_NULL(v);
+      LoDTensor* tensor_child = v->GetMutable<LoDTensor>();
+
+      ScatterUpdate<float>(dev_ctx.GetPlace(), tensor_child, &index_tensors[i],
+                           tensor_parent);
+    }
+  }
+}
+
+class CondOpProtoAndCheckerMaker : public framework::OpProtoAndCheckerMaker {
+ public:
+  CondOpProtoAndCheckerMaker(framework::OpProto* proto,
+                             framework::OpAttrChecker* op_checker)
+      : OpProtoAndCheckerMaker(proto, op_checker) {
+    AddInput("Cond", "The condition, which is a bool vector");
+    AddInput("Xs", "Inputs of Subnets").AsDuplicable();
+    AddOutput("Outs", "Outputs of Cond_Op after merge").AsDuplicable();
+
+    AddOutput("SubScopes", "sub scopes for true and false branches");
+    AddOutput("IndexTensors", "Index Tensors contains indices for true/false");
+
+    AddComment(R"DOC(
+Sample dependent Cond Operator:
+Given Cond[i] as a 1/0 vector to indicate true/false
+The equation is: 
+Out[i] = subnet_t[i], if Cond[i] == true
+Out[i] = subnet_t[i], if Cond[i] == false
+)DOC");
+  }
+};
+
+}  // namespace operators
+}  // namespace paddle
+
+REGISTER_OP_WITHOUT_GRADIENT(cond, paddle::operators::CondOp,
+                             paddle::operators::CondOpProtoAndCheckerMaker);
diff --git a/paddle/operators/cond_op.h b/paddle/operators/cond_op.h
new file mode 100644
index 0000000000000000000000000000000000000000..b09e32331e66c53555c88c06d7b1456276050eaa
--- /dev/null
+++ b/paddle/operators/cond_op.h
@@ -0,0 +1,91 @@
+/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+    http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License. */
+
+#pragma once
+#include <vector>
+#include "glog/logging.h"
+#include "paddle/framework/ddim.h"
+#include "paddle/framework/eigen.h"
+#include "paddle/framework/operator.h"
+#include "paddle/framework/tensor.h"
+#include "paddle/operators/net_op.h"
+
+namespace paddle {
+namespace operators {
+
+/*
+ * @brief CondOp is a dynamic if-else Operator
+ *
+ * It has a input tensor named cond indicating which netop each instance will
+ * run.
+ *
+ * if cond == 1, it will run true_net, which is a NetOp.
+ *
+ * if cond == 0, it will run false_net, which is another NetOp.
+ */
+class CondOp : public framework::OperatorBase {
+ public:
+  CondOp(const std::string& type, const framework::VariableNameMap& inputs,
+         const framework::VariableNameMap& outputs,
+         const framework::AttributeMap& attrs)
+      : OperatorBase(type, inputs, outputs, attrs) {
+    index_.resize(2);
+    sub_net_op_.resize(2);
+  }
+
+  CondOp(const CondOp& o)
+      : framework::OperatorBase(
+            static_cast<const framework::OperatorBase&>(o)) {
+    // TODO(yuyang18): Implement copy ctor well.
+    PADDLE_THROW("Not implemented");
+  }
+
+  void CreateScope(const framework::Scope& scope) const;
+
+  void CreateIndexTensor(const framework::Scope& scope) const;
+
+  /*
+   * InferShape must be called before Run.
+   */
+  void InferShape(const framework::Scope& scope) const override;
+
+  /*
+   * Set True Block
+   */
+  void set_truenet(std::unique_ptr<OperatorBase>&& net) {
+    sub_net_op_[0] = std::move(net);
+  }
+
+  /*
+   * Set False Block
+   */
+  void set_falsenet(std::unique_ptr<OperatorBase>&& net) {
+    sub_net_op_[1] = std::move(net);
+  }
+
+  void Run(const framework::Scope& scope,
+           const platform::DeviceContext& dev_ctx) const override;
+
+ private:
+  // sub_net_op_[0]: subnet_t
+  // sub_net_op_[1]: subnet_f
+  std::vector<std::unique_ptr<framework::OperatorBase>> sub_net_op_;
+
+  // index_[0]: True_index;
+  // index_[1]: False_index;
+  mutable std::vector<std::vector<int>> index_;
+};
+
+}  // namespace operators
+}  // namespace paddle
diff --git a/paddle/operators/cos_sim_op.cc b/paddle/operators/cos_sim_op.cc
index c033af3b741ae26ad9d37b2164f87aa6e8651c6e..72c446493684246959656dc048e7f0e761665423 100644
--- a/paddle/operators/cos_sim_op.cc
+++ b/paddle/operators/cos_sim_op.cc
@@ -25,16 +25,38 @@ class CosSimOp : public framework::OperatorWithKernel {
 
  protected:
   void InferShape(const framework::InferShapeContext &ctx) const override {
-    PADDLE_ENFORCE_NOT_NULL(ctx.InputVar("X"), "Input(X) must not be null.");
-    PADDLE_ENFORCE_NOT_NULL(ctx.InputVar("Y"), "Input(Y) must not be null.");
-    PADDLE_ENFORCE_EQ(ctx.Input<Tensor>("X")->dims(),
-                      ctx.Input<Tensor>("Y")->dims(),
-                      "Dimensions of Input(X) and Input(Y) must be the same.");
-
-    auto dims = ctx.Input<Tensor>("X")->dims();
-    ctx.Output<Tensor>("Out")->Resize({dims[0], 1});
-    ctx.Output<Tensor>("XNorm")->Resize({dims[0], 1});
-    ctx.Output<Tensor>("YNorm")->Resize({dims[0], 1});
+    // notnull check
+    PADDLE_ENFORCE_NOT_NULL(ctx.InputVar("X"),
+                            "Input(X) of CosSimOp should not be null.");
+    PADDLE_ENFORCE_NOT_NULL(ctx.InputVar("Y"),
+                            "Input(Y) of CosSimOp should not be null.");
+    PADDLE_ENFORCE_NOT_NULL(ctx.OutputVar("Out"),
+                            "Output(Out) of CosSimOp should not be null.");
+    PADDLE_ENFORCE_NOT_NULL(ctx.OutputVar("XNorm"),
+                            "Output(XNorm) of CosSimOp should not be null.");
+    PADDLE_ENFORCE_NOT_NULL(ctx.OutputVar("YNorm"),
+                            "Output(YNorm) of CosSimOp should not be null.");
+
+    // shape check
+    auto x_dims = ctx.Input<Tensor>("X")->dims();
+    auto y_dims = ctx.Input<Tensor>("Y")->dims();
+
+    PADDLE_ENFORCE_EQ(x_dims.size(), y_dims.size(),
+                      "Ranks of Input(X) and Input(Y) must be equal.");
+    PADDLE_ENFORCE_GE(x_dims.size(), 2,
+                      "Rank of Input(X) must not be less than 2.");
+    PADDLE_ENFORCE_EQ(framework::slice_ddim(x_dims, 1, x_dims.size()),
+                      framework::slice_ddim(y_dims, 1, y_dims.size()),
+                      "All dimensions except the 1st of Input(X) and Input(Y) "
+                      "must be equal.");
+    PADDLE_ENFORCE(x_dims[0] == y_dims[0] || y_dims[0] == 1,
+                   "The 1st dimension of Input(Y) must be equal to Input(X) or"
+                   " just 1 (which will be broadcasted to match Input(X)).");
+
+    // resize tensor
+    ctx.Output<framework::LoDTensor>("Out")->Resize({x_dims[0], 1});
+    ctx.Output<framework::LoDTensor>("XNorm")->Resize({x_dims[0], 1});
+    ctx.Output<framework::LoDTensor>("YNorm")->Resize({y_dims[0], 1});
   }
 };
 
@@ -42,16 +64,27 @@ class CosSimOpMaker : public framework::OpProtoAndCheckerMaker {
  public:
   CosSimOpMaker(framework::OpProto *proto, framework::OpAttrChecker *op_checker)
       : OpProtoAndCheckerMaker(proto, op_checker) {
-    AddInput("X", "The first input of cos_sim op.");
-    AddInput("Y", "The second input of cos_sim op.");
+    AddInput("X", "The 1st input of cos_sim op.");
+    AddInput("Y", "The 2nd input of cos_sim op.");
     AddOutput("Out", "The output of cos_sim op.");
-    AddOutput("XNorm", "Row norm of the first input.").AsIntermediate();
-    AddOutput("YNorm", "Row norm of the second input.").AsIntermediate();
+    AddOutput("XNorm",
+              "Norm of the first input, reduced along the 1st "
+              "dimension.")
+        .AsIntermediate();
+    AddOutput("YNorm",
+              "Norm of the second input, reduced along the 1st "
+              "dimension.")
+        .AsIntermediate();
 
     AddComment(R"DOC(
 Cosine Similarity Operator.
 
-The equation is: Out = X^T * Y / (sqrt(X^T * X) * sqrt(Y^T * Y))
+The equation is: Out = X^T * Y / (sqrt(X^T * X) * sqrt(Y^T * Y)).
+
+Input(X) and Input(Y) must have the same shape, except that the 1st dimension
+of Input(Y) could be just 1 (different from Input(X)), which will be
+broadcasted to match the shape of Input(X) before computing their cosine
+similarity.
 )DOC");
   }
 };
@@ -62,34 +95,54 @@ class CosSimOpGrad : public framework::OperatorWithKernel {
 
  protected:
   void InferShape(const framework::InferShapeContext &ctx) const override {
+    // notnull check
     PADDLE_ENFORCE_NOT_NULL(ctx.InputVar("X"), "Input(X) must not be null.");
     PADDLE_ENFORCE_NOT_NULL(ctx.InputVar("Y"), "Input(Y) must not be null.");
     PADDLE_ENFORCE_NOT_NULL(ctx.InputVar("XNorm"),
                             "Input(XNorm) must not be null.");
     PADDLE_ENFORCE_NOT_NULL(ctx.InputVar("YNorm"),
                             "Input(YNorm) must not be null.");
+    PADDLE_ENFORCE_NOT_NULL(ctx.InputVar("Out"),
+                            "Input(Out) must not be null.");
     PADDLE_ENFORCE_NOT_NULL(ctx.InputVar(framework::GradVarName("Out")),
                             "Input(Out@GRAD) must not be null.");
 
+    // shape check
     auto x_dims = ctx.Input<Tensor>("X")->dims();
     auto y_dims = ctx.Input<Tensor>("Y")->dims();
     auto xnorm_dims = ctx.Input<Tensor>("XNorm")->dims();
     auto ynorm_dims = ctx.Input<Tensor>("YNorm")->dims();
-    auto out_dims = ctx.Input<Tensor>(framework::GradVarName("Out"))->dims();
-    PADDLE_ENFORCE_EQ(x_dims, y_dims,
-                      "Dimensions of Input(X) and Input(Y) must be the same.");
-    PADDLE_ENFORCE_EQ(xnorm_dims[0], x_dims[0],
-                      "1st dimension of XNorm must equal that of Input(X).");
-    PADDLE_ENFORCE_EQ(xnorm_dims[1], 1, "2st dimension of XNorm must be one.");
-    PADDLE_ENFORCE_EQ(ynorm_dims[0], y_dims[0],
-                      "1st dimension of YNorm must equal that of Input(Y).");
-    PADDLE_ENFORCE_EQ(ynorm_dims[1], 1, "2st dimension of YNorm must be one.");
-    PADDLE_ENFORCE_EQ(out_dims[0], x_dims[0],
-                      "1st dimension of Out@GRAD must equal that of Input(X)");
-    PADDLE_ENFORCE_EQ(out_dims[1], 1, "1st dimension of Out@GRAD must be one.");
-
-    auto *x_grad = ctx.Output<Tensor>(framework::GradVarName("X"));
-    auto *y_grad = ctx.Output<Tensor>(framework::GradVarName("Y"));
+    auto out_dims = ctx.Input<Tensor>("Out")->dims();
+    auto out_grad_dims =
+        ctx.Input<Tensor>(framework::GradVarName("Out"))->dims();
+
+    PADDLE_ENFORCE_GE(x_dims.size(), y_dims.size(),
+                      "Ranks of Input(X) and Input(Y) must be equal.");
+    PADDLE_ENFORCE_GE(x_dims.size(), 2,
+                      "Rank of Input(X) must not be less than 2.");
+    PADDLE_ENFORCE_EQ(framework::slice_ddim(x_dims, 1, x_dims.size()),
+                      framework::slice_ddim(y_dims, 1, y_dims.size()),
+                      "All dimensions except the 1st of Input(X) and Input(Y) "
+                      "must be equal.");
+    PADDLE_ENFORCE(x_dims[0] == y_dims[0] || y_dims[0] == 1,
+                   "The 1st dimension of Input(Y) must be equal to Input(X) or"
+                   " just 1 (which will be broadcasted to match Input(X)).");
+    auto target_xnorm_dims = framework::make_ddim({x_dims[0], 1});
+    auto target_ynorm_dims = framework::make_ddim({y_dims[0], 1});
+    PADDLE_ENFORCE_EQ(xnorm_dims, target_xnorm_dims,
+                      "Shape of Input(XNorm) must be [X.Dim(0), 1].");
+    PADDLE_ENFORCE_EQ(ynorm_dims, target_ynorm_dims,
+                      "Shape of Input(YNorm) must be [Y.Dim(0), 1].");
+    PADDLE_ENFORCE_EQ(out_dims, target_xnorm_dims,
+                      "Shape of Input(Out) must be [X.Dim(0), 1].");
+    PADDLE_ENFORCE_EQ(out_grad_dims, target_xnorm_dims,
+                      "Shape of Input(Out@Grad) must be [X.Dim(0), 1].");
+
+    // resize tensor
+    auto *x_grad =
+        ctx.Output<framework::LoDTensor>(framework::GradVarName("X"));
+    auto *y_grad =
+        ctx.Output<framework::LoDTensor>(framework::GradVarName("Y"));
     if (x_grad) x_grad->Resize(x_dims);
     if (y_grad) y_grad->Resize(y_dims);
   }
diff --git a/paddle/operators/cos_sim_op.h b/paddle/operators/cos_sim_op.h
index 0dc509952578497671a128374f77ce616a520909..bcf6f758cae561a2e22f5be6c7a242647ef1c144 100644
--- a/paddle/operators/cos_sim_op.h
+++ b/paddle/operators/cos_sim_op.h
@@ -31,30 +31,38 @@ template <typename Place, typename T>
 class CosSimKernel : public framework::OpKernel {
  public:
   void Compute(const framework::ExecutionContext& context) const override {
-    auto* input_x = context.Input<Tensor>("X");
-    auto* input_y = context.Input<Tensor>("Y");
-    auto* output_z = context.Output<Tensor>("Out");
-    auto* output_x_norm = context.Output<Tensor>("XNorm");
-    auto* output_y_norm = context.Output<Tensor>("YNorm");
+    // get Tensor
+    auto* in_x = context.Input<Tensor>("X");
+    auto* in_y = context.Input<Tensor>("Y");
+    auto* out_z = context.Output<Tensor>("Out");
+    auto* out_x_norm = context.Output<Tensor>("XNorm");
+    auto* out_y_norm = context.Output<Tensor>("YNorm");
+    out_z->mutable_data<T>(context.GetPlace());
+    out_x_norm->mutable_data<T>(context.GetPlace());
+    out_y_norm->mutable_data<T>(context.GetPlace());
 
-    output_z->mutable_data<T>(context.GetPlace());
-    output_x_norm->mutable_data<T>(context.GetPlace());
-    output_y_norm->mutable_data<T>(context.GetPlace());
-
-    auto dims = input_x->dims();
-    int64_t size = input_x->numel();
-    auto new_dims = framework::make_ddim({dims[0], size / dims[0]});
-    auto x = EigenMatrix<T>::From(*input_x, new_dims);
-    auto y = EigenMatrix<T>::From(*input_y, new_dims);
-    auto z = EigenVector<T>::Flatten(*output_z);
-    auto x_norm = EigenVector<T>::Flatten(*output_x_norm);
-    auto y_norm = EigenVector<T>::Flatten(*output_y_norm);
+    // convert Tensor to Eigen Tensor
+    int rows_x = in_x->dims()[0];
+    int rows_y = in_y->dims()[0];
+    auto x = EigenMatrix<T>::Reshape(*in_x, 1);
+    auto y = EigenMatrix<T>::Reshape(*in_y, 1);
+    auto z = EigenVector<T>::Flatten(*out_z);
+    auto x_norm = EigenVector<T>::Flatten(*out_x_norm);
+    auto y_norm = EigenVector<T>::Flatten(*out_y_norm);
 
+    // compute
     auto place = context.GetEigenDevice<Place>();
-    auto xy = (x * y).sum(Eigen::array<int, 1>({{1}}));
-    x_norm.device(place) = x.square().sum(Eigen::array<int, 1>({{1}})).sqrt();
-    y_norm.device(place) = y.square().sum(Eigen::array<int, 1>({{1}})).sqrt();
-    z.device(place) = xy / x_norm / y_norm;
+    auto row_along = Eigen::array<int, 1>({{1}});
+    x_norm.device(place) = x.square().sum(row_along).sqrt();
+    y_norm.device(place) = y.square().sum(row_along).sqrt();
+    if (rows_x == rows_y) {
+      auto xy = (x * y).sum(Eigen::array<int, 1>({{1}}));
+      z.device(place) = xy / x_norm / y_norm;
+    } else {
+      Eigen::DSizes<int, 2> bcast(rows_x, 1);
+      auto xy = (x * y.broadcast(bcast)).sum(row_along);
+      z.device(place) = xy / x_norm / y_norm.broadcast(bcast);
+    }
   }
 };
 
@@ -62,43 +70,72 @@ template <typename Place, typename T>
 class CosSimGradKernel : public framework::OpKernel {
  public:
   void Compute(const framework::ExecutionContext& context) const override {
-    auto* input_x = context.Input<Tensor>("X");
-    auto* input_y = context.Input<Tensor>("Y");
-    auto* input_z = context.Input<Tensor>("Out");
-    auto* input_x_norm = context.Input<Tensor>("XNorm");
-    auto* input_y_norm = context.Input<Tensor>("YNorm");
-    auto* output_grad_x = context.Output<Tensor>(framework::GradVarName("X"));
-    auto* output_grad_y = context.Output<Tensor>(framework::GradVarName("Y"));
-    auto* input_grad_z = context.Input<Tensor>(framework::GradVarName("Out"));
+    // get Tensor
+    auto* in_x = context.Input<Tensor>("X");
+    auto* in_y = context.Input<Tensor>("Y");
+    auto* in_z = context.Input<Tensor>("Out");
+    auto* in_x_norm = context.Input<Tensor>("XNorm");
+    auto* in_y_norm = context.Input<Tensor>("YNorm");
+    auto* out_grad_x = context.Output<Tensor>(framework::GradVarName("X"));
+    auto* out_grad_y = context.Output<Tensor>(framework::GradVarName("Y"));
+    auto* in_grad_z = context.Input<Tensor>(framework::GradVarName("Out"));
 
-    auto dims = input_x->dims();
-    int64_t size = input_x->numel();
-    auto new_dims = framework::make_ddim({dims[0], size / dims[0]});
-    auto x = EigenMatrix<T>::From(*input_x, new_dims);
-    auto y = EigenMatrix<T>::From(*input_y, new_dims);
-    auto z = EigenMatrix<T>::From(*input_z);
-    auto x_norm = EigenMatrix<T>::From(*input_x_norm);
-    auto y_norm = EigenMatrix<T>::From(*input_y_norm);
-    auto dz = EigenMatrix<T>::From(*input_grad_z);
+    // convert Tensor to Eigen Tensor
+    auto x = EigenMatrix<T>::Reshape(*in_x, 1);
+    auto y = EigenMatrix<T>::Reshape(*in_y, 1);
+    auto z = EigenMatrix<T>::Reshape(*in_z, 1);
+    auto x_norm = EigenMatrix<T>::Reshape(*in_x_norm, 1);
+    auto y_norm = EigenMatrix<T>::Reshape(*in_y_norm, 1);
+    auto dz = EigenMatrix<T>::Reshape(*in_grad_z, 1);
 
-    Eigen::DSizes<int, 2> bcast(1, new_dims[1]);
-    auto z_bcast = z.broadcast(bcast);
-    auto dz_bcast = dz.broadcast(bcast);
+    // compute gradident
+    int rows_x = in_x->dims()[0];
+    int rows_y = in_y->dims()[0];
+    int cols = framework::product(in_x->dims()) / rows_x;
+    Eigen::DSizes<int, 2> bcast_cols(1, cols);
+    auto z_bcast = z.broadcast(bcast_cols);
+    auto dz_bcast = dz.broadcast(bcast_cols);
+    auto x_snorm_bcast = x_norm.square().eval().broadcast(bcast_cols);
     auto place = context.GetEigenDevice<Place>();
-    auto x_snorm_bcast = x_norm.square().eval().broadcast(bcast);
-    auto y_snorm_bcast = y_norm.square().eval().broadcast(bcast);
-    auto norm_prod_bcast = (x_norm * y_norm).eval().broadcast(bcast);
-    if (output_grad_x) {
-      output_grad_x->mutable_data<T>(context.GetPlace());
-      auto dx = EigenMatrix<T>::From(*output_grad_x, new_dims);
-      dx.device(place) =
-          dz_bcast * (y / norm_prod_bcast - z_bcast * x / x_snorm_bcast);
-    }
-    if (output_grad_y) {
-      output_grad_y->mutable_data<T>(context.GetPlace());
-      auto dy = EigenMatrix<T>::From(*output_grad_y, new_dims);
-      dy.device(place) =
-          dz_bcast * (x / norm_prod_bcast - z_bcast * y / y_snorm_bcast);
+    if (rows_x == rows_y) {
+      auto y_snorm_bcast = y_norm.square().eval().broadcast(bcast_cols);
+      auto norm_prod_bcast = (x_norm * y_norm).eval().broadcast(bcast_cols);
+      // compute dx
+      if (out_grad_x) {
+        out_grad_x->mutable_data<T>(context.GetPlace());
+        auto dx = EigenMatrix<T>::Reshape(*out_grad_x, 1);
+        auto grad = y / norm_prod_bcast - z_bcast * x / x_snorm_bcast;
+        dx.device(place) = dz_bcast * grad;
+      }
+      // compute dy
+      if (out_grad_y) {
+        out_grad_y->mutable_data<T>(context.GetPlace());
+        auto dy = EigenMatrix<T>::Reshape(*out_grad_y, 1);
+        auto grad = x / norm_prod_bcast - z_bcast * y / y_snorm_bcast;
+        dy.device(place) = dz_bcast * grad;
+      }
+    } else {
+      Eigen::DSizes<int, 2> bcast_rows(rows_x, 1);
+      Eigen::DSizes<int, 2> bcast_rows_cols(rows_x, cols);
+      auto y_bcast = y.broadcast(bcast_rows);
+      auto y_snorm_bcast = y_norm.square().eval().broadcast(bcast_rows_cols);
+      auto norm_prod_bcast = (x_norm * y_norm.eval().broadcast(bcast_rows))
+                                 .eval()
+                                 .broadcast(bcast_cols);
+      // compute dx
+      if (out_grad_x) {
+        out_grad_x->mutable_data<T>(context.GetPlace());
+        auto dx = EigenMatrix<T>::Reshape(*out_grad_x, 1);
+        auto grad = y_bcast / norm_prod_bcast - z_bcast * x / x_snorm_bcast;
+        dx.device(place) = dz_bcast * grad;
+      }
+      // compute dy
+      if (out_grad_y) {
+        out_grad_y->mutable_data<T>(context.GetPlace());
+        auto dy = EigenMatrix<T>::Reshape(*out_grad_y, 1);
+        auto grad = x / norm_prod_bcast - z_bcast * y_bcast / y_snorm_bcast;
+        dy.device(place) = (dz_bcast * grad).sum(Eigen::array<int, 1>({{0}}));
+      }
     }
   }
 };
diff --git a/paddle/operators/cross_entropy_op.cc b/paddle/operators/cross_entropy_op.cc
index 32ad0e82fa4612553ad14854432f4162adcb7587..c31c13289851b2bd86a4fb92417f04b5c5b3a515 100644
--- a/paddle/operators/cross_entropy_op.cc
+++ b/paddle/operators/cross_entropy_op.cc
@@ -17,12 +17,21 @@ limitations under the License. */
 namespace paddle {
 namespace operators {
 
+using framework::LoDTensor;
+
 class CrossEntropyOp : public framework::OperatorWithKernel {
  public:
   using framework::OperatorWithKernel::OperatorWithKernel;
 
  protected:
   void InferShape(const framework::InferShapeContext &ctx) const override {
+    PADDLE_ENFORCE_NOT_NULL(ctx.InputVar("X"),
+                            "Input(X) of CrossEntropyOp must not be null.");
+    PADDLE_ENFORCE_NOT_NULL(ctx.InputVar("Label"),
+                            "Input(Label) of CrossEntropyOp must not be null.");
+    PADDLE_ENFORCE_NOT_NULL(ctx.OutputVar("Y"),
+                            "Output(Y) of CrossEntropyOp must not be null.");
+
     auto *x = ctx.Input<Tensor>("X");
     auto *label = ctx.Input<Tensor>("Label");
 
@@ -35,7 +44,7 @@ class CrossEntropyOp : public framework::OperatorWithKernel {
       // normal cross entropy
       PADDLE_ENFORCE_EQ(x->dims()[0], label->dims()[0]);
     }
-    ctx.Output<Tensor>("Y")->Resize({x->dims()[0]});
+    ctx.Output<LoDTensor>("Y")->Resize({x->dims()[0], 1});
   }
 };
 
@@ -45,7 +54,7 @@ class CrossEntropyGradientOp : public framework::OperatorWithKernel {
 
  protected:
   void InferShape(const framework::InferShapeContext &ctx) const override {
-    auto dx = ctx.Output<Tensor>(framework::GradVarName("X"));
+    auto dx = ctx.Output<LoDTensor>(framework::GradVarName("X"));
     auto x = ctx.Input<Tensor>("X");
 
     dx->Resize(x->dims());
diff --git a/paddle/operators/elementwise_mul_op.cc b/paddle/operators/elementwise_mul_op.cc
new file mode 100644
index 0000000000000000000000000000000000000000..ee6e975b443691bf71cec904565ced20406f3fba
--- /dev/null
+++ b/paddle/operators/elementwise_mul_op.cc
@@ -0,0 +1,117 @@
+/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve.
+
+   Licensed under the Apache License, Version 2.0 (the "License");
+   you may not use this file except in compliance with the License.
+   You may obtain a copy of the License at
+
+   http://www.apache.org/licenses/LICENSE-2.0
+
+   Unless required by applicable law or agreed to in writing, software
+   distributed under the License is distributed on an "AS IS" BASIS,
+   WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+   See the License for the specific language governing permissions and
+   limitations under the License. */
+
+#include "paddle/operators/elementwise_mul_op.h"
+
+namespace paddle {
+namespace operators {
+
+using Tensor = framework::Tensor;
+
+class ElementWiseMulOp : public framework::OperatorWithKernel {
+ public:
+  using framework::OperatorWithKernel::OperatorWithKernel;
+
+ protected:
+  void InferShape(const framework::InferShapeContext &ctx) const override {
+    PADDLE_ENFORCE_NOT_NULL(ctx.InputVar("X"),
+                            "Input(X) of ElementWiseMulOp should not be null.");
+    PADDLE_ENFORCE_NOT_NULL(ctx.InputVar("Y"),
+                            "Input(Y) of ElementWiseMulOp should not be null.");
+    PADDLE_ENFORCE_NOT_NULL(
+        ctx.OutputVar("Out"),
+        "Output(Out) of ElementWiseMulOp should not be null.");
+
+    auto x_dim = ctx.Input<Tensor>("X")->dims();
+    auto y_dim = ctx.Input<Tensor>("Y")->dims();
+    PADDLE_ENFORCE_GE(x_dim.size(), y_dim.size(),
+                      "Rank of first input must >= rank of second input.")
+    ctx.Output<framework::LoDTensor>("Out")->Resize(x_dim);
+  }
+};
+
+class ElementWiseMulOpMaker : public framework::OpProtoAndCheckerMaker {
+ public:
+  ElementWiseMulOpMaker(framework::OpProto *proto,
+                        framework::OpAttrChecker *op_checker)
+      : OpProtoAndCheckerMaker(proto, op_checker) {
+    AddInput("X", "The first input of elementwise mul op");
+    AddInput("Y", "The second input of elementwise mul op");
+    AddAttr<int>("axis",
+                 R"DOC(
+When shape(Y) does not equal shape(X),Y will be broadcasted 
+to match the shape of X and axis should be dimension index Y in X
+        )DOC")
+        .SetDefault(-1)
+        .EqualGreaterThan(-1);
+
+    AddOutput("Out", "The output of elementwise mul op");
+    AddComment(R"DOC(
+Limited elementwise multiple operator.The equation is: Out = X ⊙ Y.
+1. The shape of Y should be same with X or
+2. Y's shape is a subset of X. 
+   Y will be broadcasted to match the shape of X and axis should be dimension index Y in X.
+   example:
+      shape(X) = (2, 3, 4, 5), shape(Y) = (,)
+      shape(X) = (2, 3, 4, 5), shape(Y) = (5,)
+      shape(X) = (2, 3, 4, 5), shape(Y) = (4, 5)
+      shape(X) = (2, 3, 4, 5), shape(Y) = (3, 4), with axis=1
+      shape(X) = (2, 3, 4, 5), shape(Y) = (2), with axis=0
+)DOC");
+  }
+};
+
+class ElementWiseMulOpGrad : public framework::OperatorWithKernel {
+ public:
+  using framework::OperatorWithKernel::OperatorWithKernel;
+
+ protected:
+  void InferShape(const framework::InferShapeContext &ctx) const override {
+    PADDLE_ENFORCE_NOT_NULL(ctx.InputVar("X"), "Input(X) should not be null");
+    PADDLE_ENFORCE_NOT_NULL(ctx.InputVar("Y"), "Input(Y) should not be null");
+    PADDLE_ENFORCE_NOT_NULL(ctx.InputVar(framework::GradVarName("Out")),
+                            "Input(Out@GRAD) should not be null");
+
+    auto x_dims = ctx.Input<Tensor>("X")->dims();
+    auto y_dims = ctx.Input<Tensor>("Y")->dims();
+    auto out_dims = ctx.Input<Tensor>(framework::GradVarName("Out"))->dims();
+    auto *x_grad =
+        ctx.Output<framework::LoDTensor>(framework::GradVarName("X"));
+    auto *y_grad =
+        ctx.Output<framework::LoDTensor>(framework::GradVarName("Y"));
+
+    PADDLE_ENFORCE_GE(x_dims.size(), y_dims.size(),
+                      "Rank of first input must >= rank of second input.")
+
+    if (x_grad) {
+      x_grad->Resize(x_dims);
+    }
+
+    if (y_grad) {
+      y_grad->Resize(y_dims);
+    }
+  }
+};
+}  // namespace operators
+}  // namespace paddle
+
+namespace ops = paddle::operators;
+REGISTER_OP(elementwise_mul, ops::ElementWiseMulOp, ops::ElementWiseMulOpMaker,
+            elementwise_mul_grad, ops::ElementWiseMulOpGrad);
+REGISTER_OP_CPU_KERNEL(
+    elementwise_mul,
+    ops::ElementWiseMulKernel<paddle::platform::CPUPlace, float>);
+REGISTER_OP_CPU_KERNEL(
+    elementwise_mul_grad,
+    ops::ElementWiseMulGradKernel<paddle::platform::CPUPlace, float>);
diff --git a/paddle/operators/elementwise_mul_op.cu b/paddle/operators/elementwise_mul_op.cu
new file mode 100644
index 0000000000000000000000000000000000000000..56f2087c22c6c599a3c5aef36eb0fe3eac295bef
--- /dev/null
+++ b/paddle/operators/elementwise_mul_op.cu
@@ -0,0 +1,25 @@
+/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve.
+
+   Licensed under the Apache License, Version 2.0 (the "License");
+   you may not use this file except in compliance with the License.
+   You may obtain a copy of the License at
+
+   http://www.apache.org/licenses/LICENSE-2.0
+
+   Unless required by applicable law or agreed to in writing, software
+   distributed under the License is distributed on an "AS IS" BASIS,
+   WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+   See the License for the specific language governing permissions and
+   limitations under the License. */
+
+#define EIGEN_USE_GPU
+#include "paddle/operators/elementwise_mul_op.h"
+
+namespace ops = paddle::operators;
+
+REGISTER_OP_GPU_KERNEL(
+    elementwise_mul,
+    ops::ElementWiseMulKernel<paddle::platform::GPUPlace, float>);
+REGISTER_OP_GPU_KERNEL(
+    elementwise_mul_grad,
+    ops::ElementWiseMulGradKernel<paddle::platform::GPUPlace, float>);
diff --git a/paddle/operators/elementwise_mul_op.h b/paddle/operators/elementwise_mul_op.h
new file mode 100644
index 0000000000000000000000000000000000000000..6d58da580b81b9e0a8ae170eec1a73638b190df8
--- /dev/null
+++ b/paddle/operators/elementwise_mul_op.h
@@ -0,0 +1,183 @@
+/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve.
+
+   Licensed under the Apache License, Version 2.0 (the "License");
+   you may not use this file except in compliance with the License.
+   You may obtain a copy of the License at
+
+   http://www.apache.org/licenses/LICENSE-2.0
+
+   Unless required by applicable law or agreed to in writing, software
+   distributed under the License is distributed on an "AS IS" BASIS,
+   WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+   See the License for the specific language governing permissions and
+   limitations under the License. */
+
+#pragma once
+#include "paddle/framework/eigen.h"
+#include "paddle/framework/op_registry.h"
+
+namespace paddle {
+namespace operators {
+/*
+ * Out = X ⊙ Y
+ * 1. shape(X) = (2, 3, 4, 5), shape(Y) = (3, 4), with axis=1
+ *    pre=2, n=3*4, post=5
+ * 2. shape(X) = (2, 3, 4, 5), shape(Y) = (4,5)
+ *    pre=2*3, n=4*5, post=1
+ */
+
+inline void get_mid_dims(const framework::DDim& x_dims,
+                         const framework::DDim& y_dims, const int axis,
+                         int& pre, int& n, int& post) {
+  pre = 1;
+  n = 1;
+  post = 1;
+  for (int i = 0; i < axis; ++i) {
+    pre *= x_dims[i];
+  }
+
+  for (int i = 0; i < y_dims.size(); ++i) {
+    PADDLE_ENFORCE_EQ(x_dims[i + axis], y_dims[i],
+                      "Broadcast dimension mismatch.");
+    n *= y_dims[i];
+  }
+
+  for (int i = axis + y_dims.size(); i < x_dims.size(); ++i) {
+    post *= x_dims[i];
+  }
+}
+
+template <typename Place, typename T>
+class ElementWiseMulKernel : public framework::OpKernel {
+ public:
+  void Compute(const framework::ExecutionContext& ctx) const override {
+    using Tensor = framework::Tensor;
+
+    auto* x = ctx.Input<Tensor>("X");
+    auto* y = ctx.Input<Tensor>("Y");
+    auto* z = ctx.Output<Tensor>("Out");
+    z->mutable_data<T>(ctx.GetPlace());
+
+    auto x_e = framework::EigenVector<T>::Flatten(*x);
+    auto y_e = framework::EigenVector<T>::Flatten(*y);
+    auto z_e = framework::EigenVector<T>::Flatten(*z);
+
+    auto x_dims = x->dims();
+    auto y_dims = y->dims();
+    PADDLE_ENFORCE_GE(x_dims.size(), y_dims.size(),
+                      "Rank of first input must >= rank of second input.")
+
+    if (x_dims == y_dims || product(y_dims) == 1) {
+      z_e.device(ctx.GetEigenDevice<Place>()) = x_e * y_e;
+      return;
+    }
+
+    int axis = ctx.Attr<int>("axis");
+    axis = (axis == -1 ? x_dims.size() - y_dims.size() : axis);
+    PADDLE_ENFORCE(axis >= 0 && axis < x_dims.size(),
+                   "Axis should be in range [0, x_dims)");
+
+    int pre, n, post;
+    get_mid_dims(x_dims, y_dims, axis, pre, n, post);
+    if (post == 1) {
+      auto y_bcast = y_e.reshape(Eigen::DSizes<int, 2>(1, n))
+                         .broadcast(Eigen::DSizes<int, 2>(pre, 1))
+                         .reshape(Eigen::DSizes<int, 1>(x_e.size()));
+      z_e.device(ctx.GetEigenDevice<Place>()) = x_e * y_bcast;
+      return;
+    } else {
+      auto y_bcast = y_e.reshape(Eigen::DSizes<int, 3>(1, n, 1))
+                         .broadcast(Eigen::DSizes<int, 3>(pre, 1, post))
+                         .reshape(Eigen::DSizes<int, 1>(x_e.size()));
+      z_e.device(ctx.GetEigenDevice<Place>()) = x_e * y_bcast;
+      return;
+    }
+  }
+};
+
+template <typename Place, typename T>
+class ElementWiseMulGradKernel : public framework::OpKernel {
+ public:
+  void Compute(const framework::ExecutionContext& ctx) const override {
+    using Tensor = framework::Tensor;
+
+    auto* x = ctx.Input<Tensor>("X");
+    auto* y = ctx.Input<Tensor>("Y");
+    auto* dout = ctx.Input<Tensor>(framework::GradVarName("Out"));
+
+    auto x_e = framework::EigenVector<T>::Flatten(*x);
+    auto y_e = framework::EigenVector<T>::Flatten(*y);
+    auto dout_e = framework::EigenVector<T>::Flatten(*dout);
+
+    auto x_dims = x->dims();
+    auto y_dims = y->dims();
+
+    auto* dx = ctx.Output<Tensor>(framework::GradVarName("X"));
+    auto* dy = ctx.Output<Tensor>(framework::GradVarName("Y"));
+    if (dx) {
+      dx->mutable_data<T>(ctx.GetPlace());
+    }
+    if (dy) {
+      dy->mutable_data<T>(ctx.GetPlace());
+    }
+
+    if (x_dims == y_dims || product(y_dims) == 1) {
+      if (dx) {
+        auto dx_e = framework::EigenVector<T>::Flatten(*dx);
+        dx_e.device(ctx.GetEigenDevice<Place>()) = dout_e * y_e;
+      }
+
+      if (dy) {
+        auto dy_e = framework::EigenVector<T>::Flatten(*dy);
+        dy_e.device(ctx.GetEigenDevice<Place>()) = x_e * dout_e;
+      }
+      return;
+    }
+
+    int axis = ctx.Attr<int>("axis");
+    axis = (axis == -1 ? x_dims.size() - y_dims.size() : axis);
+
+    int pre, n, post;
+    get_mid_dims(x_dims, y_dims, axis, pre, n, post);
+
+    // TODO(gongweibao): wrap reshape to a function.
+    if (post == 1) {
+      auto y_e_bcast = y_e.reshape(Eigen::DSizes<int, 2>(1, n))
+                           .broadcast(Eigen::DSizes<int, 2>(pre, 1))
+                           .reshape(Eigen::DSizes<int, 1>(x_e.size()));
+      if (dx) {
+        auto dx_e = framework::EigenVector<T>::Flatten(*dx);
+        dx_e.device(ctx.GetEigenDevice<Place>()) = dout_e * y_e_bcast;
+      }
+
+      if (dy) {
+        auto dy_e = framework::EigenVector<T>::Flatten(*dy);
+        dy_e.device(ctx.GetEigenDevice<Place>()) =
+            (x_e * dout_e)
+                .reshape(Eigen::DSizes<int, 2>(pre, n))
+                .sum(Eigen::array<int, 1>{{0}});
+      }
+      return;
+    } else {
+      auto y_e_bcast = y_e.reshape(Eigen::DSizes<int, 3>(1, n, 1))
+                           .broadcast(Eigen::DSizes<int, 3>(pre, 1, post))
+                           .reshape(Eigen::DSizes<int, 1>(x_e.size()));
+      if (dx) {
+        auto dx_e = framework::EigenVector<T>::Flatten(*dx);
+        dx_e.device(ctx.GetEigenDevice<Place>()) = dout_e * y_e_bcast;
+      }
+
+      if (dy) {
+        auto dy_e = framework::EigenVector<T>::Flatten(*dy);
+        dy_e.device(ctx.GetEigenDevice<Place>()) =
+            (x_e * dout_e)
+                .reshape(Eigen::DSizes<int, 3>(pre, n, post))
+                .sum(Eigen::array<int, 2>{{0, 2}});
+      }
+      return;
+    }
+  }
+};
+
+}  // namespace operators
+}  // namespace paddle
diff --git a/paddle/operators/fill_zeros_like_op.cc b/paddle/operators/fill_zeros_like_op.cc
index 9d51f6e3a16fe96125599bb440d40237aeb9a028..ba7857cc65f6860a6156674c6addc2bfdce21a99 100644
--- a/paddle/operators/fill_zeros_like_op.cc
+++ b/paddle/operators/fill_zeros_like_op.cc
@@ -23,7 +23,14 @@ class FillZerosLikeOp : public framework::OperatorWithKernel {
 
  protected:
   void InferShape(const framework::InferShapeContext &ctx) const override {
-    ctx.Output<framework::Tensor>("Dst")->Resize(
+    PADDLE_ENFORCE_NOT_NULL(
+        ctx.InputVar("Src"),
+        "Input(Src) of FillZerosLikeOp should not be null.");
+    PADDLE_ENFORCE_NOT_NULL(
+        ctx.OutputVar("Dst"),
+        "Output(Dst) of FillZerosLikeOp should not be null.");
+
+    ctx.Output<framework::LoDTensor>("Dst")->Resize(
         ctx.Input<framework::Tensor>("Src")->dims());
   }
 };
diff --git a/paddle/operators/gather_op.cc b/paddle/operators/gather_op.cc
index 123bed296c462c30bddd3bfbd530098fdbfe4856..d445b61c1657356f2cdcf1e98d756607de2bd042 100644
--- a/paddle/operators/gather_op.cc
+++ b/paddle/operators/gather_op.cc
@@ -24,11 +24,18 @@ class GatherOp : public framework::OperatorWithKernel {
 
  protected:
   void InferShape(const framework::InferShapeContext &ctx) const override {
+    PADDLE_ENFORCE_NOT_NULL(ctx.InputVar("X"),
+                            "Input(X) of GatherOp should not be null.");
+    PADDLE_ENFORCE_NOT_NULL(ctx.InputVar("Index"),
+                            "Input(Index) of GatherOp should not be null.");
+    PADDLE_ENFORCE_NOT_NULL(ctx.OutputVar("Out"),
+                            "Output(Out) of GatherOp should not be null.");
+
     int batch_size = ctx.Input<Tensor>("Index")->dims()[0];
     PADDLE_ENFORCE_GE(batch_size, 0, "Batch size must be >0");
     framework::DDim output_dims(ctx.Input<Tensor>("X")->dims());
     output_dims[0] = batch_size;
-    ctx.Output<Tensor>("Out")->Resize(output_dims);
+    ctx.Output<framework::LoDTensor>("Out")->Resize(output_dims);
   }
 };
 
@@ -38,7 +45,7 @@ class GatherGradOp : public framework::OperatorWithKernel {
 
  protected:
   void InferShape(const framework::InferShapeContext &ctx) const override {
-    auto X_grad = ctx.Output<Tensor>(framework::GradVarName("X"));
+    auto X_grad = ctx.Output<framework::LoDTensor>(framework::GradVarName("X"));
     auto X = ctx.Input<Tensor>("X");
 
     X_grad->Resize(X->dims());
diff --git a/paddle/operators/gaussian_random_op.cc b/paddle/operators/gaussian_random_op.cc
index 3d76516405960c502a46997108049b2db5cab6bf..c0e161bbc0c5486eb10408e43e6388f1b287abf8 100644
--- a/paddle/operators/gaussian_random_op.cc
+++ b/paddle/operators/gaussian_random_op.cc
@@ -43,8 +43,12 @@ class GaussianRandomOp : public framework::OperatorWithKernel {
   using framework::OperatorWithKernel::OperatorWithKernel;
 
  protected:
-  void InferShape(const framework::InferShapeContext& context) const override {
-    auto* tensor = context.Output<framework::Tensor>("Out");
+  void InferShape(const framework::InferShapeContext& ctx) const override {
+    PADDLE_ENFORCE_NOT_NULL(
+        ctx.OutputVar("Out"),
+        "Output(Out) of GaussianRandomOp should not be null.");
+
+    auto* tensor = ctx.Output<framework::LoDTensor>("Out");
     auto dims = Attr<std::vector<int>>("dims");
     std::vector<int64_t> temp;
     temp.reserve(dims.size());
diff --git a/paddle/operators/identity_op.cc b/paddle/operators/identity_op.cc
index 7d9d4fa519d1c690feacbadc5175aeab49082282..b67ca5f6f8d516224e18a5eed497f2bfc680259c 100644
--- a/paddle/operators/identity_op.cc
+++ b/paddle/operators/identity_op.cc
@@ -42,6 +42,11 @@ class IdentityOp : public NetOp {
              const framework::VariableNameMap &outputs,
              const framework::AttributeMap &attrs)
       : NetOp(type, inputs, outputs, attrs) {
+    PADDLE_ENFORCE_NE(Input("X"), framework::kEmptyVarName,
+                      "Input(X) of IdentityOp should not be null.");
+    PADDLE_ENFORCE_NE(Output("Out"), framework::kEmptyVarName,
+                      "Output(Out) of IdentityOp should not be null.");
+
     AppendOp(framework::OpRegistry::CreateOp(
         "scale", {{"X", {Input("X")}}}, {{"Out", {Output("Out")}}},
         {{"scale", static_cast<AttrType>(1)}}));
diff --git a/paddle/operators/lookup_table_op.cc b/paddle/operators/lookup_table_op.cc
index 94d40890a765413e88a35a6ad995ca97ac84dcda..07f6dfabca5879e3de6004e59d2e87f7fa68d66c 100644
--- a/paddle/operators/lookup_table_op.cc
+++ b/paddle/operators/lookup_table_op.cc
@@ -22,10 +22,17 @@ class LookupTableOp : public framework::OperatorWithKernel {
   using framework::OperatorWithKernel::OperatorWithKernel;
 
  protected:
-  void InferShape(const framework::InferShapeContext &context) const override {
-    auto table_t = context.Input<Tensor>("W");
-    auto ids_t = context.Input<Tensor>("Ids");
-    auto output_t = context.Output<Tensor>("Out");
+  void InferShape(const framework::InferShapeContext &ctx) const override {
+    PADDLE_ENFORCE_NOT_NULL(ctx.InputVar("W"),
+                            "Input(W) of LookupTableOp should not be null.");
+    PADDLE_ENFORCE_NOT_NULL(ctx.InputVar("Ids"),
+                            "Input(Ids) of LookupTableOp should not be null.");
+    PADDLE_ENFORCE_NOT_NULL(ctx.OutputVar("Out"),
+                            "Output(Out) of LookupTableOp should not be null.");
+
+    auto table_t = ctx.Input<Tensor>("W");
+    auto ids_t = ctx.Input<Tensor>("Ids");
+    auto output_t = ctx.Output<framework::LoDTensor>("Out");
 
     output_t->Resize({ids_t->dims()[0], table_t->dims()[1]});
   }
@@ -56,7 +63,8 @@ class LookupTableOpGrad : public framework::OperatorWithKernel {
  protected:
   void InferShape(const framework::InferShapeContext &context) const override {
     auto table = context.Input<Tensor>("W");
-    auto d_table = context.Output<Tensor>(framework::GradVarName("W"));
+    auto d_table =
+        context.Output<framework::LoDTensor>(framework::GradVarName("W"));
     d_table->Resize(table->dims());
   }
 };
diff --git a/paddle/operators/mean_op.cc b/paddle/operators/mean_op.cc
index d3d0e55a674587fb04f43f24d0790de4358f035a..7d7eeb59a23435036dc33c1e4fe6dd1c4a1a2f62 100644
--- a/paddle/operators/mean_op.cc
+++ b/paddle/operators/mean_op.cc
@@ -24,8 +24,10 @@ class MeanOp : public framework::OperatorWithKernel {
  protected:
   void InferShape(const framework::InferShapeContext &ctx) const override {
     PADDLE_ENFORCE_NOT_NULL(ctx.InputVar("X"),
-                            "Input of MeanOp must be initialized.");
-    ctx.Output<Tensor>("Out")->Resize({1});
+                            "Input(X) of MeanOp should not be null.");
+    PADDLE_ENFORCE_NOT_NULL(ctx.OutputVar("Out"),
+                            "Output(Out) of MeanOp should not be null.");
+    ctx.Output<framework::LoDTensor>("Out")->Resize({1});
   }
 };
 
@@ -45,7 +47,7 @@ class MeanGradOp : public framework::OperatorWithKernel {
 
  protected:
   void InferShape(const framework::InferShapeContext &ctx) const override {
-    ctx.Output<Tensor>(framework::GradVarName("X"))
+    ctx.Output<framework::LoDTensor>(framework::GradVarName("X"))
         ->Resize(ctx.Input<Tensor>("X")->dims());
   }
 };
diff --git a/paddle/operators/minus_op.cc b/paddle/operators/minus_op.cc
index a4876feb2edf77bd422fa2a7687b0fa7d55dae47..ecf8a6f7795314e2475bb9546b55b8f354b96366 100644
--- a/paddle/operators/minus_op.cc
+++ b/paddle/operators/minus_op.cc
@@ -27,13 +27,20 @@ class MinusOp : public framework::OperatorWithKernel {
 
  protected:
   void InferShape(const framework::InferShapeContext &ctx) const override {
+    PADDLE_ENFORCE_NOT_NULL(ctx.InputVar("X"),
+                            "Input(X) of MinusOp should not be null.");
+    PADDLE_ENFORCE_NOT_NULL(ctx.InputVar("Y"),
+                            "Input(Y) of MinusOp should not be null.");
+    PADDLE_ENFORCE_NOT_NULL(ctx.OutputVar("Out"),
+                            "Output(Out) of MinusOp should not be null.");
+
     auto *left_tensor = ctx.Input<framework::Tensor>("X");
     auto *right_tensor = ctx.Input<framework::Tensor>("Y");
 
     PADDLE_ENFORCE_EQ(
         left_tensor->numel(), right_tensor->numel(),
         "Minus operator must take two tensor with same num of elements");
-    ctx.Output<framework::Tensor>("Out")->Resize(left_tensor->dims());
+    ctx.Output<framework::LoDTensor>("Out")->Resize(left_tensor->dims());
   }
 };
 
@@ -77,8 +84,6 @@ class MinusGradOp : public NetOp {
 }  // namespace operators
 }  // namespace paddle
 
-USE_OP(scale);
-USE_NO_KERNEL_OP(identity);
 namespace ops = paddle::operators;
 REGISTER_OP(minus, ops::MinusOp, ops::MinusOpMaker, minus_grad,
             ops::MinusGradOp<float>);
diff --git a/paddle/operators/mul_op.cc b/paddle/operators/mul_op.cc
index 710a56a0e8e2d17162d7d000df226f1537104eb9..b6d320b415e02549e85cb36ab517b0b5433887d5 100644
--- a/paddle/operators/mul_op.cc
+++ b/paddle/operators/mul_op.cc
@@ -18,6 +18,7 @@ namespace paddle {
 namespace operators {
 
 using framework::Tensor;
+using framework::LoDTensor;
 
 class MulOp : public framework::OperatorWithKernel {
  public:
@@ -25,6 +26,13 @@ class MulOp : public framework::OperatorWithKernel {
 
  protected:
   void InferShape(const framework::InferShapeContext &ctx) const override {
+    PADDLE_ENFORCE_NOT_NULL(ctx.InputVar("X"),
+                            "Input(X) of MulOp should not be null.");
+    PADDLE_ENFORCE_NOT_NULL(ctx.InputVar("Y"),
+                            "Input(Y) of MulOp should not be null.");
+    PADDLE_ENFORCE_NOT_NULL(ctx.OutputVar("Out"),
+                            "Output(Out) of MulOp should not be null.");
+
     auto x_dims = ctx.Input<Tensor>("X")->dims();
     auto y_dims = ctx.Input<Tensor>("Y")->dims();
     int x_num_col_dims = Attr<int>("x_num_col_dims");
@@ -45,7 +53,8 @@ class MulOp : public framework::OperatorWithKernel {
     PADDLE_ENFORCE_EQ(
         x_mat_dims[1], y_mat_dims[0],
         "First matrix's width must be equal with second matrix's height.");
-    ctx.Output<Tensor>("Out")->Resize({x_mat_dims[0], y_mat_dims[1]});
+    ctx.Output<framework::LoDTensor>("Out")->Resize(
+        {x_mat_dims[0], y_mat_dims[1]});
   }
 };
 
@@ -94,8 +103,10 @@ class MulOpGrad : public framework::OperatorWithKernel {
     auto x_dims = ctx.Input<Tensor>("X")->dims();
     auto y_dims = ctx.Input<Tensor>("Y")->dims();
     auto out_dims = ctx.Input<Tensor>(framework::GradVarName("Out"))->dims();
-    auto *x_grad = ctx.Output<Tensor>(framework::GradVarName("X"));
-    auto *y_grad = ctx.Output<Tensor>(framework::GradVarName("Y"));
+    auto *x_grad =
+        ctx.Output<framework::LoDTensor>(framework::GradVarName("X"));
+    auto *y_grad =
+        ctx.Output<framework::LoDTensor>(framework::GradVarName("Y"));
 
     auto x_mat_dims =
         framework::flatten_to_2d(x_dims, Attr<int>("x_num_col_dims"));
diff --git a/paddle/operators/name_convention.md b/paddle/operators/name_convention.md
index a090e0b5450509affdd739f63df618595f204f97..379385dc5d914101c7b5c9494f9383b6cf6a9b79 100644
--- a/paddle/operators/name_convention.md
+++ b/paddle/operators/name_convention.md
@@ -38,9 +38,11 @@ public:
   AccumulateOpMaker(framework::OpProto *proto,
                             framework::OpAttrChecker *op_checker)
     : OpProtoAndCheckerMaker(proto, op_checker) {
-    AddInput("X", "(Tensor) The input tensor that has to be accumulated to the output tensor. If the output size is not the same as input size, the output tensor is first reshaped and initialized to zero, and only then, accumulation is done.");
+    AddInput("X", "(Tensor) The input tensor that has to be accumulated to the output tensor. 
+    If the output size is not the same as input size, 
+    the output tensor is first reshaped and initialized to zero, and only then, accumulation is done.");
     AddOutput("Out", "(Tensor) Accumulated output tensor");
-    AddAttr<float>("gamma", "(float, default 1.0) Accumulation multiplier");
+    AddAttr<float>("gamma", "(float, default 1.0) Accumulation multiplier").SetDefault(1.0f);
     AddComment(R"DOC(
 Accumulate operator accumulates the input tensor to the output tensor. If the
 output tensor already has the right size, we add to it; otherwise, we first
@@ -51,7 +53,7 @@ Accumulation is done as shown:
 
 Out = 1*X + gamma*Out
 
-where X is the input tensor, Y is the output tensor and gamma is the multiplier
+where X is the input tensor, Out is the output tensor and gamma is the multiplier
 argument.
 )DOC");
   }
diff --git a/paddle/operators/pad_op.cc b/paddle/operators/pad_op.cc
new file mode 100644
index 0000000000000000000000000000000000000000..a0b1c6b631d97a40d774f7d2ff9550fda9c32db4
--- /dev/null
+++ b/paddle/operators/pad_op.cc
@@ -0,0 +1,118 @@
+/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve.
+
+   Licensed under the Apache License, Version 2.0 (the "License");
+   you may not use this file except in compliance with the License.
+   You may obtain a copy of the License at
+
+   http://www.apache.org/licenses/LICENSE-2.0
+
+   Unless required by applicable law or agreed to in writing, software
+   distributed under the License is distributed on an "AS IS" BASIS,
+   WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+   See the License for the specific language governing permissions and
+   limitations under the License. */
+
+#include "paddle/operators/pad_op.h"
+
+namespace paddle {
+namespace operators {
+
+using framework::Tensor;
+
+class PadOp : public framework::OperatorWithKernel {
+ public:
+  using framework::OperatorWithKernel::OperatorWithKernel;
+
+ protected:
+  void InferShape(const framework::InferShapeContext &ctx) const override {
+    PADDLE_ENFORCE_NOT_NULL(ctx.InputVar("X"),
+                            "Input(X) of PadOp should not be null.");
+    PADDLE_ENFORCE_NOT_NULL(ctx.OutputVar("Out"),
+                            "Output(Out) of PadOp should not be null.");
+
+    auto x_dim = ctx.Input<Tensor>("X")->dims();
+    auto paddings = Attr<std::vector<int>>("paddings");
+    PADDLE_ENFORCE_EQ(x_dim.size() * 2, int64_t(paddings.size()),
+                      "Size of paddings should be equal to 2 * dimension size "
+                      "of input tensor.");
+    std::vector<int64_t> out_dims(x_dim.size());
+    for (int i = 0; i < x_dim.size(); ++i) {
+      out_dims[i] = x_dim[i] + paddings[i * 2] + paddings[i * 2 + 1];
+    }
+    ctx.Output<framework::LoDTensor>("Out")->Resize(
+        framework::make_ddim(out_dims));
+  }
+};
+
+class PadOpMaker : public framework::OpProtoAndCheckerMaker {
+ public:
+  PadOpMaker(framework::OpProto *proto, framework::OpAttrChecker *op_checker)
+      : OpProtoAndCheckerMaker(proto, op_checker) {
+    AddInput("X",
+             "The input of pad op. "
+             "The input should be a k-D tensor(k > 0 and k < 7)");
+    AddOutput("Out",
+              "The output of pad op."
+              "A tensor with the same shape as X.")
+        .NotInGradient();
+    AddComment(R"DOC(
+Pad input into output, as specified by paddings and pad_value. The input should be a k-D tensor(k > 0 and k < 7). As an example:
+
+Given:
+
+X = [[1, 2],
+   [3, 4]]
+
+and 
+
+paddings = [0, 1, 1, 2]
+
+and
+ 
+pad_value = 0 
+
+then we get 
+
+Out = [[0, 1, 2, 0, 0]
+       [0, 3, 4, 0, 0]
+       [0, 0, 0, 0, 0]]
+)DOC");
+    AddAttr<std::vector<int>>(
+        "paddings",
+        "A list<int> to describes padding rules for each dimension."
+        " For 2-D image tensor, paddings=[0, 1, 2, 3] means"
+        " padding 0 row to top, 1 row to bottom, 2 columns to left"
+        " and 3 columns to right.Size of paddings should be equal to"
+        " 2 * dimension size of input tensor.");
+    AddAttr<float>("pad_value",
+                   "(float) default to 0; "
+                   "The value to fill padded areas.")
+        .SetDefault(0.0f);
+  }
+};
+
+class PadOpGrad : public framework::OperatorWithKernel {
+ public:
+  using framework::OperatorWithKernel::OperatorWithKernel;
+
+ protected:
+  void InferShape(const framework::InferShapeContext &ctx) const override {
+    PADDLE_ENFORCE_NOT_NULL(ctx.InputVar("X"), "Input(X) should not be null");
+    PADDLE_ENFORCE_NOT_NULL(ctx.InputVar(framework::GradVarName("Out")),
+                            "Input(Out@GRAD) should not be null");
+    auto x_dims = ctx.Input<Tensor>("X")->dims();
+    auto *x_g = ctx.Output<framework::LoDTensor>(framework::GradVarName("X"));
+    if (x_g != nullptr) {
+      x_g->Resize(x_dims);
+    }
+  }
+};
+
+}  // namespace operators
+}  // namespace paddle
+
+namespace ops = paddle::operators;
+REGISTER_OP(pad, ops::PadOp, ops::PadOpMaker, pad_grad, ops::PadOpGrad);
+REGISTER_OP_CPU_KERNEL(pad, ops::PadKernel<paddle::platform::CPUPlace, float>);
+REGISTER_OP_CPU_KERNEL(pad_grad,
+                       ops::PadGradKernel<paddle::platform::CPUPlace, float>);
diff --git a/paddle/operators/pad_op.cu b/paddle/operators/pad_op.cu
new file mode 100644
index 0000000000000000000000000000000000000000..555a7dba23c6fa2659cabf4858b42ff70d74bf18
--- /dev/null
+++ b/paddle/operators/pad_op.cu
@@ -0,0 +1,21 @@
+/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve.
+
+   Licensed under the Apache License, Version 2.0 (the "License");
+   you may not use this file except in compliance with the License.
+   You may obtain a copy of the License at
+
+   http://www.apache.org/licenses/LICENSE-2.0
+
+   Unless required by applicable law or agreed to in writing, software
+   distributed under the License is distributed on an "AS IS" BASIS,
+   WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+   See the License for the specific language governing permissions and
+   limitations under the License. */
+
+#define EIGEN_USE_GPU
+#include "paddle/operators/pad_op.h"
+
+namespace ops = paddle::operators;
+REGISTER_OP_GPU_KERNEL(pad, ops::PadKernel<paddle::platform::GPUPlace, float>);
+REGISTER_OP_GPU_KERNEL(pad_grad,
+                       ops::PadGradKernel<paddle::platform::GPUPlace, float>);
diff --git a/paddle/operators/pad_op.h b/paddle/operators/pad_op.h
new file mode 100644
index 0000000000000000000000000000000000000000..2cc3b945ae5b2e2e93d8531c7f99e4c215d1d806
--- /dev/null
+++ b/paddle/operators/pad_op.h
@@ -0,0 +1,132 @@
+/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve.
+
+   Licensed under the Apache License, Version 2.0 (the "License");
+   you may not use this file except in compliance with the License.
+   You may obtain a copy of the License at
+
+   http://www.apache.org/licenses/LICENSE-2.0
+
+   Unless required by applicable law or agreed to in writing, software
+   distributed under the License is distributed on an "AS IS" BASIS,
+   WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+   See the License for the specific language governing permissions and
+   limitations under the License. */
+
+#pragma once
+
+#include "paddle/framework/eigen.h"
+#include "paddle/framework/op_registry.h"
+
+namespace paddle {
+namespace operators {
+
+using Tensor = framework::Tensor;
+
+template <typename T, size_t D, int MajorType = Eigen::RowMajor,
+          typename IndexType = Eigen::DenseIndex>
+using EigenTensor = framework::EigenTensor<T, D, MajorType, IndexType>;
+
+template <typename Place, typename T, size_t D>
+void PadFunction(const framework::ExecutionContext& context) {
+  auto pads = context.Attr<std::vector<int>>("paddings");
+  Eigen::array<std::pair<int, int>, D> paddings;
+  for (size_t i = 0; i < paddings.size(); ++i) {
+    paddings[i].first = pads[i * 2];
+    paddings[i].second = pads[i * 2 + 1];
+  }
+  T pad_value = context.Attr<T>("pad_value");
+
+  auto* x = context.Input<Tensor>("X");
+  auto* out = context.Output<Tensor>("Out");
+  out->mutable_data<T>(context.GetPlace());
+
+  auto x_tensor = EigenTensor<T, D>::From(*x);
+  auto out_tensor = EigenTensor<T, D>::From(*out);
+  auto place = context.GetEigenDevice<Place>();
+  out_tensor.device(place) = x_tensor.pad(paddings, pad_value);
+}
+
+template <typename Place, typename T>
+class PadKernel : public framework::OpKernel {
+ public:
+  void Compute(const framework::ExecutionContext& context) const override {
+    int rank = context.Input<Tensor>("X")->dims().size();
+    switch (rank) {
+      case 1:
+        PadFunction<Place, T, 1>(context);
+        break;
+      case 2:
+        PadFunction<Place, T, 2>(context);
+        break;
+      case 3:
+        PadFunction<Place, T, 3>(context);
+        break;
+      case 4:
+        PadFunction<Place, T, 4>(context);
+        break;
+      case 5:
+        PadFunction<Place, T, 5>(context);
+        break;
+      case 6:
+        PadFunction<Place, T, 6>(context);
+        break;
+      default:
+        PADDLE_THROW(
+            "PadOp only support tensors with no more than 6 dimensions.");
+    }
+  }
+};
+
+template <typename Place, typename T, size_t D>
+void PadGradFunction(const framework::ExecutionContext& context) {
+  auto pads = context.Attr<std::vector<int>>("paddings");
+  Eigen::array<std::pair<int, int>, D> paddings;
+  for (size_t i = 0; i < paddings.size(); ++i) {
+    paddings[i].first = -pads[i * 2];
+    paddings[i].second = -pads[i * 2 + 1];
+  }
+  auto* d_out = context.Input<Tensor>(framework::GradVarName("Out"));
+  auto* d_x = context.Output<Tensor>(framework::GradVarName("X"));
+  if (d_x != nullptr) {
+    d_x->mutable_data<T>(context.GetPlace());
+    auto d_x_tensor = EigenTensor<T, D>::From(*d_x);
+    auto d_out_tensor = EigenTensor<T, D>::From(*d_out);
+    auto place = context.GetEigenDevice<Place>();
+    d_x_tensor.device(place) = d_out_tensor.pad(paddings, 0);
+  }
+}
+
+template <typename Place, typename T>
+class PadGradKernel : public framework::OpKernel {
+ public:
+  void Compute(const framework::ExecutionContext& context) const override {
+    size_t rank =
+        context.Input<Tensor>(framework::GradVarName("Out"))->dims().size();
+    switch (rank) {
+      case 1:
+        PadGradFunction<Place, T, 1>(context);
+        break;
+      case 2:
+        PadGradFunction<Place, T, 2>(context);
+        break;
+      case 3:
+        PadGradFunction<Place, T, 3>(context);
+        break;
+      case 4:
+        PadGradFunction<Place, T, 4>(context);
+        break;
+      case 5:
+        PadGradFunction<Place, T, 5>(context);
+        break;
+      case 6:
+        PadGradFunction<Place, T, 6>(context);
+        break;
+      default:
+        PADDLE_THROW(
+            "PadOp only support tensors with no more than 6 dimensions.");
+    }
+  }
+};
+
+}  // namespace operators
+}  // namespace paddle
diff --git a/paddle/operators/recurrent_op.cc b/paddle/operators/recurrent_op.cc
index e826703c60ca82e1fe690eb78c3d4f92981ef3a2..d3413d7cb9305732e9ddf3cb1bc267f7203097f3 100644
--- a/paddle/operators/recurrent_op.cc
+++ b/paddle/operators/recurrent_op.cc
@@ -26,10 +26,11 @@ namespace operators {
 using Scope = framework::Scope;
 using Variable = framework::Variable;
 using Tensor = framework::Tensor;
+using LoDTensor = framework::LoDTensor;
 
 void RecurrentAlgorithm::InferShape(const Scope& scope) const {
   seq_len_ = scope.FindVar((arg_->inlinks[0]).external)
-                 ->GetMutable<Tensor>()
+                 ->GetMutable<LoDTensor>()
                  ->dims()[0];
   CreateScopes(scope);
   auto step_scopes = GetStepScopes(scope);
@@ -88,7 +89,7 @@ void RecurrentAlgorithm::CreateScopes(const Scope& scope) const {
         // the weight are located in parent scope
         for (auto& var_name : input.second) {
           if (!step_scope.FindVar(var_name)) {
-            step_scope.NewVar(var_name)->GetMutable<Tensor>();
+            step_scope.NewVar(var_name)->GetMutable<LoDTensor>();
           }
         }
       }
@@ -106,11 +107,12 @@ void RecurrentAlgorithm::CreateScopes(const Scope& scope) const {
 void RecurrentAlgorithm::InitMemories(Scope* step_scope,
                                       bool infer_shape_mode) const {
   for (auto& attr : arg_->memories) {
-    Tensor* pre_mem = step_scope->NewVar(attr.pre_var)->GetMutable<Tensor>();
+    auto* pre_mem = step_scope->NewVar(attr.pre_var)->GetMutable<LoDTensor>();
     PADDLE_ENFORCE(step_scope->FindVar(attr.boot_var) != nullptr,
                    "memory [%s]'s boot variable [%s] not exists", attr.var,
                    attr.boot_var);
-    Tensor* boot_mem = step_scope->FindVar(attr.boot_var)->GetMutable<Tensor>();
+    auto* boot_mem =
+        step_scope->FindVar(attr.boot_var)->GetMutable<LoDTensor>();
     if (infer_shape_mode) {
       pre_mem->Resize(boot_mem->dims());
       PADDLE_ENFORCE_EQ(pre_mem->dims().size(), 2);
@@ -192,9 +194,9 @@ void RecurrentGradientAlgorithm::LinkBootMemoryGradients(
                    "memory variable [%s] does not exists", attr.var);
     PADDLE_ENFORCE(step_scope->FindVar(attr.boot_var) != nullptr,
                    "boot variable [%s] does not exists", attr.boot_var);
-    Tensor* mem_grad = step_scope->NewVar(attr.var)->GetMutable<Tensor>();
-    Tensor* boot_mem_grad =
-        step_scope->NewVar(attr.boot_var)->GetMutable<Tensor>();
+    auto* mem_grad = step_scope->NewVar(attr.var)->GetMutable<LoDTensor>();
+    auto* boot_mem_grad =
+        step_scope->NewVar(attr.boot_var)->GetMutable<LoDTensor>();
     if (infer_shape_mode) {
       boot_mem_grad->Resize(mem_grad->dims());
     } else {
@@ -205,7 +207,7 @@ void RecurrentGradientAlgorithm::LinkBootMemoryGradients(
 
 void RecurrentGradientAlgorithm::InferShape(const Scope& scope) const {
   seq_len_ = scope.FindVar((arg_->inlinks[0]).external)
-                 ->GetMutable<Tensor>()
+                 ->GetMutable<LoDTensor>()
                  ->dims()[0];
   auto step_scopes = GetStepScopes(scope);
   rnn::SegmentInputs(step_scopes, arg_->inlinks, seq_len_,
diff --git a/paddle/operators/reshape_op.cc b/paddle/operators/reshape_op.cc
index b7061153d2bf13982f14f233e87a87daeeebf5fd..0d05e344148c68f5625dd819ec59c5991892e4ce 100644
--- a/paddle/operators/reshape_op.cc
+++ b/paddle/operators/reshape_op.cc
@@ -28,7 +28,11 @@ class ReshapeOp : public framework::OperatorWithKernel {
  protected:
   void InferShape(const framework::InferShapeContext &ctx) const override {
     // input check
-    PADDLE_ENFORCE_NOT_NULL(ctx.InputVar("X"), "Input(X) shouldn't be null");
+    PADDLE_ENFORCE_NOT_NULL(ctx.InputVar("X"),
+                            "Input(X) of ReshapeOp should not be null.");
+    PADDLE_ENFORCE_NOT_NULL(ctx.OutputVar("Out"),
+                            "Output(Out) of ReshapeOp should not be null.");
+
     auto shape = ctx.Attr<std::vector<int>>("shape");
     PADDLE_ENFORCE(shape.size() > 0, "Attr(shape) shouldn't be empty.");
     for (auto dim : shape) {
@@ -46,7 +50,7 @@ class ReshapeOp : public framework::OperatorWithKernel {
     std::transform(shape.begin(), shape.end(), shape_int64.begin(),
                    [](int a) { return static_cast<int64_t>(a); });
     auto out_dims = framework::make_ddim(shape_int64);
-    ctx.Output<framework::Tensor>("Out")->Resize(out_dims);
+    ctx.Output<framework::LoDTensor>("Out")->Resize(out_dims);
   }
 };
 
@@ -90,7 +94,7 @@ class ReshapeGradOp : public framework::OperatorWithKernel {
     PADDLE_ENFORCE_NOT_NULL(ctx.InputVar(framework::GradVarName("Out")),
                             "Input(Out@GRAD) shouldn't be null.");
     auto dims = ctx.Input<framework::Tensor>("X")->dims();
-    auto *d_in = ctx.Output<framework::Tensor>(framework::GradVarName("X"));
+    auto *d_in = ctx.Output<framework::LoDTensor>(framework::GradVarName("X"));
     d_in->Resize(dims);
   }
 };
diff --git a/paddle/operators/rnn/recurrent_op_utils.cc b/paddle/operators/rnn/recurrent_op_utils.cc
index 97872c67ac99fbf6c9c177d52f1d4069163e8548..6c082cb1825e04accb09019fef28eb2ec6523a5b 100644
--- a/paddle/operators/rnn/recurrent_op_utils.cc
+++ b/paddle/operators/rnn/recurrent_op_utils.cc
@@ -21,6 +21,7 @@ namespace rnn {
 namespace f = paddle::framework;
 
 using Tensor = framework::Tensor;
+using LoDTensor = framework::LoDTensor;
 
 void SegmentInputs(const std::vector<Scope*>& step_scopes,
                    const std::vector<Link>& inlinks, const size_t seq_len,
@@ -31,7 +32,7 @@ void SegmentInputs(const std::vector<Scope*>& step_scopes,
     PADDLE_ENFORCE(input_var != nullptr, "input link [%s] is not in scope.",
                    inlinks[i].external);
 
-    Tensor* input = input_var->GetMutable<Tensor>();
+    LoDTensor* input = input_var->GetMutable<LoDTensor>();
     f::DDim dims = input->dims();
     PADDLE_ENFORCE(static_cast<size_t>(dims[0]) == seq_len,
                    "all the inlinks must have same length");
@@ -40,6 +41,8 @@ void SegmentInputs(const std::vector<Scope*>& step_scopes,
       Tensor* step_input =
           step_scopes[j]->NewVar(inlinks[i].internal)->GetMutable<Tensor>();
       if (!infer_shape_mode) {
+        // The input of operators of each step is Tensor here.
+        // Maybe need to modify Slice function.
         *step_input = input->Slice<float>(j, j + 1);
       }
       step_input->Resize(step_dims);
@@ -54,21 +57,23 @@ void ConcatOutputs(const std::vector<Scope*>& step_scopes,
     auto output_var = step_scopes[0]->FindVar(outlinks[i].external);
     PADDLE_ENFORCE(output_var != nullptr, "output link [%s] is not in scope.",
                    outlinks[i].external);
-    Tensor* output = output_var->GetMutable<Tensor>();
+    LoDTensor* output = output_var->GetMutable<LoDTensor>();
 
     if (infer_shape_mode) {
       auto step_scope_var = step_scopes[0]->FindVar(outlinks[i].internal);
       PADDLE_ENFORCE(step_scope_var != nullptr, "%s not in scope",
                      outlinks[i].internal);
-      f::DDim step_dims = step_scope_var->template GetMutable<Tensor>()->dims();
+      f::DDim step_dims =
+          step_scope_var->template GetMutable<LoDTensor>()->dims();
       std::vector<int64_t> dims_vec = vectorize(step_dims);
       dims_vec.insert(dims_vec.begin(), seq_len);
       output->Resize(f::make_ddim(dims_vec));
     } else {
       output->mutable_data<float>(platform::CPUPlace());
       for (size_t j = 0; j < seq_len; j++) {
-        Tensor* step_output =
-            step_scopes[j]->FindVar(outlinks[i].internal)->GetMutable<Tensor>();
+        LoDTensor* step_output = step_scopes[j]
+                                     ->FindVar(outlinks[i].internal)
+                                     ->GetMutable<LoDTensor>();
         // TODO(luotao02) data type and platform::DeviceContext() should set
         // correctly
         (output->Slice<float>(j, j + 1))
@@ -94,8 +99,8 @@ void LinkMemories(const std::vector<Scope*>& scopes,
   auto scope = scopes[step_id];
   auto linked_scope = scopes[step_id + offset];
   for (auto& attr : memories) {
-    auto mem = scope->FindVar(attr.pre_var)->GetMutable<Tensor>();
-    auto linked_mem = linked_scope->FindVar(attr.var)->GetMutable<Tensor>();
+    auto mem = scope->FindVar(attr.pre_var)->GetMutable<LoDTensor>();
+    auto linked_mem = linked_scope->FindVar(attr.var)->GetMutable<LoDTensor>();
     if (infer_shape_mode) {
       mem->Resize(linked_mem->dims());
     } else {
diff --git a/paddle/operators/rowwise_add_op.cc b/paddle/operators/rowwise_add_op.cc
index fa8f0ff1a858143af427b51025279c726f1628e0..2a3fd3be941d91aaa6b014df91d3025f07767577 100644
--- a/paddle/operators/rowwise_add_op.cc
+++ b/paddle/operators/rowwise_add_op.cc
@@ -25,6 +25,13 @@ class RowwiseAddOp : public framework::OperatorWithKernel {
 
  protected:
   void InferShape(const framework::InferShapeContext &ctx) const override {
+    PADDLE_ENFORCE_NOT_NULL(ctx.InputVar("X"),
+                            "Input(X) of RowwiseAddOp should not be null.");
+    PADDLE_ENFORCE_NOT_NULL(ctx.InputVar("b"),
+                            "Input(b) of RowwiseAddOp should not be null.");
+    PADDLE_ENFORCE_NOT_NULL(ctx.OutputVar("Out"),
+                            "Output(Out) of RowwiseAddOp should not be null.");
+
     auto x_dims = ctx.Input<Tensor>("X")->dims();
     auto b_dims = ctx.Input<Tensor>("b")->dims();
     PADDLE_ENFORCE_GT(
@@ -37,7 +44,7 @@ class RowwiseAddOp : public framework::OperatorWithKernel {
         framework::slice_ddim(x_dims, num_col_dims, x_dims.size()), b_dims,
         "The width of two operands must be same");
     PADDLE_ENFORCE_EQ(ctx.OutputSize("Out"), 1, "The output size must be 1");
-    ctx.Output<Tensor>("Out")->Resize(x_dims);
+    ctx.Output<framework::LoDTensor>("Out")->Resize(x_dims);
   }
 };
 
@@ -76,8 +83,8 @@ class RowwiseAddGradOp : public framework::OperatorWithKernel {
     PADDLE_ENFORCE_EQ(
         framework::slice_ddim(x_dims, num_col_dims, x_dims.size()), b_dims,
         "The width of two operands must be same");
-    auto *dx = ctx.Output<Tensor>(framework::GradVarName("X"));
-    auto *db = ctx.Output<Tensor>(framework::GradVarName("b"));
+    auto *dx = ctx.Output<framework::LoDTensor>(framework::GradVarName("X"));
+    auto *db = ctx.Output<framework::LoDTensor>(framework::GradVarName("b"));
     if (dx) dx->Resize(x_dims);
     if (db) db->Resize(b_dims);
   }
diff --git a/paddle/operators/scale_op.cc b/paddle/operators/scale_op.cc
index ea991f683d841b3dc4624a0d8aa3c88367fd3c6d..d1f42e8662537d35e17429f9d436fdc0e5a1dc11 100644
--- a/paddle/operators/scale_op.cc
+++ b/paddle/operators/scale_op.cc
@@ -27,8 +27,13 @@ class ScaleOp : public framework::OperatorWithKernel {
 
  protected:
   void InferShape(const framework::InferShapeContext &ctx) const override {
+    PADDLE_ENFORCE_NOT_NULL(ctx.InputVar("X"),
+                            "Input(X) of ScaleOp should not be null.");
+    PADDLE_ENFORCE_NOT_NULL(ctx.OutputVar("Out"),
+                            "Output(Out) of ScaleOp should not be null.");
+
     auto *in = ctx.Input<framework::Tensor>("X");
-    auto *out = ctx.Output<framework::Tensor>("Out");
+    auto *out = ctx.Output<framework::LoDTensor>("Out");
     out->Resize(in->dims());
   }
 };
diff --git a/paddle/operators/scatter_op.cc b/paddle/operators/scatter_op.cc
index f901edefa22dc9a252e87116df756d04767a7162..8820262732327306f4f807702751708bd1e2aa36 100644
--- a/paddle/operators/scatter_op.cc
+++ b/paddle/operators/scatter_op.cc
@@ -24,6 +24,15 @@ class ScatterOp : public framework::OperatorWithKernel {
 
  protected:
   void InferShape(const framework::InferShapeContext &ctx) const override {
+    PADDLE_ENFORCE_NOT_NULL(ctx.InputVar("Ref"),
+                            "Input(Ref) of ScatterOp should not be null.");
+    PADDLE_ENFORCE_NOT_NULL(ctx.InputVar("Index"),
+                            "Input(Index) of ScatterOp should not be null.");
+    PADDLE_ENFORCE_NOT_NULL(ctx.InputVar("Updates"),
+                            "Input(Updates) of ScatterOp should not be null.");
+    PADDLE_ENFORCE_NOT_NULL(ctx.OutputVar("Out"),
+                            "Output(Out) of ScatterOp should not be null.");
+
     PADDLE_ENFORCE_EQ(ctx.Input<Tensor>("Index")->dims().size(), 1,
                       "Update Index should be 1-D.");
     PADDLE_ENFORCE_EQ(ctx.Input<Tensor>("Ref")->dims().size(),
@@ -35,7 +44,8 @@ class ScatterOp : public framework::OperatorWithKernel {
     framework::DDim data_dim(ctx.Input<Tensor>("Updates")->dims());
     for (int i = 1; i < data_dim.size(); ++i)
       PADDLE_ENFORCE_EQ(data_dim[i], ctx.Input<Tensor>("Updates")->dims()[i]);
-    ctx.Output<Tensor>("Out")->Resize(ctx.Input<Tensor>("Ref")->dims());
+    ctx.Output<framework::LoDTensor>("Out")->Resize(
+        ctx.Input<Tensor>("Ref")->dims());
   }
 };
 
@@ -45,9 +55,11 @@ class ScatterGradOp : public framework::OperatorWithKernel {
 
  protected:
   void InferShape(const framework::InferShapeContext &ctx) const override {
-    auto *dUpdates = ctx.Output<Tensor>(framework::GradVarName("Updates"));
+    auto *dUpdates =
+        ctx.Output<framework::LoDTensor>(framework::GradVarName("Updates"));
     auto *Updates = ctx.Input<Tensor>("Updates");
-    auto *dRef = ctx.Output<Tensor>(framework::GradVarName("Ref"));
+    auto *dRef =
+        ctx.Output<framework::LoDTensor>(framework::GradVarName("Ref"));
     auto *Ref = ctx.Input<Tensor>("Ref");
 
     dRef->Resize(Ref->dims());
diff --git a/paddle/operators/sequence_avg_pool_op.cc b/paddle/operators/sequence_avg_pool_op.cc
new file mode 100644
index 0000000000000000000000000000000000000000..9815b8f3a8d813959949bbfedc79f404721a8216
--- /dev/null
+++ b/paddle/operators/sequence_avg_pool_op.cc
@@ -0,0 +1,95 @@
+/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+    http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License. */
+
+#include "paddle/operators/sequence_avg_pool_op.h"
+
+namespace paddle {
+namespace operators {
+
+class SequenceAvgPoolOp : public framework::OperatorWithKernel {
+ public:
+  using framework::OperatorWithKernel::OperatorWithKernel;
+
+ protected:
+  void InferShape(const framework::InferShapeContext& ctx) const override {
+    PADDLE_ENFORCE_NOT_NULL(
+        ctx.InputVar("X"), "Input(X) of SequenceAvgPoolOp should not be null.");
+    PADDLE_ENFORCE_NOT_NULL(
+        ctx.OutputVar("Out"),
+        "Output(Out) of SequenceAvgPoolOp should not be null.");
+
+    auto* x = ctx.Input<framework::LoDTensor>("X");
+    auto dims = x->dims();
+    auto lod = x->lod();
+    PADDLE_ENFORCE_EQ(lod.size(), 1UL, "Only support one level sequence now.");
+    PADDLE_ENFORCE_GE(
+        dims[0],
+        /*batch size = */ static_cast<int64_t>(lod[0].size() - 1),
+        "The first dimension of Input(X) must be large than batch size.");
+    dims[0] = lod[0].size() - 1;
+    ctx.Output<framework::LoDTensor>("Out")->Resize({dims});
+  }
+};
+
+class SequenceAvgPoolOpMaker : public framework::OpProtoAndCheckerMaker {
+ public:
+  SequenceAvgPoolOpMaker(framework::OpProto* proto,
+                         framework::OpAttrChecker* op_checker)
+      : OpProtoAndCheckerMaker(proto, op_checker) {
+    AddInput("X", "Input of SequenceAvgPoolOp.");
+    AddOutput("Out", "The output of SequenceAvgPoolOp.");
+    AddComment(R"DOC(
+    SequenceAvgPoolOp averages features of all time-steps of each instance.
+    More detailed comments will be added later.
+    )DOC");
+  }
+};
+
+class SequenceAvgPoolGradOp : public framework::OperatorWithKernel {
+ public:
+  using framework::OperatorWithKernel::OperatorWithKernel;
+
+ protected:
+  void InferShape(const framework::InferShapeContext& ctx) const override {
+    PADDLE_ENFORCE_NOT_NULL(ctx.InputVar(framework::GradVarName("Out")),
+                            "Gradient of Out should not be null.");
+    PADDLE_ENFORCE_NOT_NULL(ctx.InputVar("X"),
+                            "The input X should not be null.");
+    auto og_dims =
+        ctx.Input<framework::LoDTensor>(framework::GradVarName("Out"))->dims();
+    auto x_dims = ctx.Input<framework::LoDTensor>("X")->dims();
+    PADDLE_ENFORCE_EQ(og_dims.size(), x_dims.size(),
+                      "The rank of output grad must equal to Input(X).");
+    for (int64_t i = 1; i < og_dims.size(); ++i) {
+      PADDLE_ENFORCE_EQ(og_dims[i], x_dims[i], "The dimension mismatch.");
+    }
+    auto* x_grad =
+        ctx.Output<framework::LoDTensor>(framework::GradVarName("X"));
+    x_grad->Resize(x_dims);
+  }
+};
+
+}  // namespace operators
+}  // namespace paddle
+
+namespace ops = paddle::operators;
+REGISTER_OP(sequence_avg_pool, ops::SequenceAvgPoolOp,
+            ops::SequenceAvgPoolOpMaker, sequence_avg_pool_grad,
+            ops::SequenceAvgPoolGradOp);
+REGISTER_OP_CPU_KERNEL(
+    sequence_avg_pool,
+    ops::SequenceAvgPoolKernel<paddle::platform::CPUPlace, float>);
+REGISTER_OP_CPU_KERNEL(
+    sequence_avg_pool_grad,
+    ops::SequenceAvgPoolGradKernel<paddle::platform::CPUPlace, float>);
diff --git a/paddle/operators/sequence_avg_pool_op.cu b/paddle/operators/sequence_avg_pool_op.cu
new file mode 100644
index 0000000000000000000000000000000000000000..bc9d1611fccd17c99b914b6ef59995288a9ebbd6
--- /dev/null
+++ b/paddle/operators/sequence_avg_pool_op.cu
@@ -0,0 +1,25 @@
+/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve.
+
+   Licensed under the Apache License, Version 2.0 (the "License");
+   you may not use this file except in compliance with the License.
+   You may obtain a copy of the License at
+
+   http://www.apache.org/licenses/LICENSE-2.0
+
+   Unless required by applicable law or agreed to in writing, software
+   distributed under the License is distributed on an "AS IS" BASIS,
+   WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+   See the License for the specific language governing permissions and
+   limitations under the License. */
+
+#define EIGEN_USE_GPU
+
+#include "paddle/operators/sequence_avg_pool_op.h"
+
+namespace ops = paddle::operators;
+REGISTER_OP_GPU_KERNEL(
+    sequence_avg_pool,
+    ops::SequenceAvgPoolKernel<paddle::platform::GPUPlace, float>);
+REGISTER_OP_GPU_KERNEL(
+    sequence_avg_pool_grad,
+    ops::SequenceAvgPoolGradKernel<paddle::platform::GPUPlace, float>);
diff --git a/paddle/operators/sequence_avg_pool_op.h b/paddle/operators/sequence_avg_pool_op.h
new file mode 100644
index 0000000000000000000000000000000000000000..ebe0956344eb71d0fb2836f1b4a989ac546d9f78
--- /dev/null
+++ b/paddle/operators/sequence_avg_pool_op.h
@@ -0,0 +1,84 @@
+/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+    http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License. */
+
+#pragma once
+#include "paddle/framework/eigen.h"
+#include "paddle/framework/op_registry.h"
+
+namespace paddle {
+namespace operators {
+
+using Tensor = framework::Tensor;
+using LoDTensor = framework::LoDTensor;
+template <typename T, int MajorType = Eigen::RowMajor,
+          typename IndexType = Eigen::DenseIndex>
+using EigenVector = framework::EigenVector<T, MajorType, IndexType>;
+template <typename T, int MajorType = Eigen::RowMajor,
+          typename IndexType = Eigen::DenseIndex>
+using EigenMatrix = framework::EigenMatrix<T, MajorType, IndexType>;
+
+template <typename Place, typename T>
+class SequenceAvgPoolKernel : public framework::OpKernel {
+ public:
+  void Compute(const framework::ExecutionContext& context) const override {
+    auto* in = context.Input<LoDTensor>("X");
+    auto* out = context.Output<LoDTensor>("Out");
+
+    auto dims = in->dims();
+    auto lod = in->lod();
+    int64_t w = in->numel() / dims[0];
+
+    out->mutable_data<T>(context.GetPlace());
+    auto place = context.GetEigenDevice<Place>();
+    for (int i = 0; i < static_cast<int>(lod[0].size()) - 1; ++i) {
+      Tensor in_t = in->Slice<T>(static_cast<int>(lod[0][i]),
+                                 static_cast<int>(lod[0][i + 1]));
+      Tensor out_t = out->Slice<T>(i, i + 1);
+      int64_t h = static_cast<int64_t>(lod[0][i + 1] - lod[0][i]);
+      auto in_e = EigenMatrix<T>::From(in_t, framework::make_ddim({h, w}));
+      auto out_e = EigenVector<T>::Flatten(out_t);
+      out_e.device(place) = in_e.mean(Eigen::array<int, 1>({{0}}));
+    }
+  }
+};
+
+template <typename Place, typename T>
+class SequenceAvgPoolGradKernel : public framework::OpKernel {
+ public:
+  void Compute(const framework::ExecutionContext& context) const override {
+    auto* in = context.Input<LoDTensor>("X");
+    auto* out_g = context.Input<LoDTensor>(framework::GradVarName("Out"));
+    auto* in_g = context.Output<LoDTensor>(framework::GradVarName("X"));
+
+    auto dims = in->dims();
+    auto lod = in->lod();
+    int64_t w = in->numel() / dims[0];
+
+    in_g->mutable_data<T>(context.GetPlace());
+    auto place = context.GetEigenDevice<Place>();
+    for (int i = 0; i < static_cast<int>(lod[0].size()) - 1; ++i) {
+      auto in_g_t = in_g->Slice<T>(static_cast<int>(lod[0][i]),
+                                   static_cast<int>(lod[0][i + 1]));
+      auto out_g_t = out_g->Slice<T>(i, i + 1);
+      int64_t h = static_cast<int64_t>(lod[0][i + 1] - lod[0][i]);
+      auto in_g_e = EigenMatrix<T>::From(in_g_t, {h, w});
+      auto out_g_e = EigenMatrix<T>::From(out_g_t, {1, w});
+      Eigen::DSizes<int, 2> bcast(h, 1);
+      in_g_e.device(place) = (out_g_e / static_cast<T>(h)).broadcast(bcast);
+    }
+  }
+};
+
+}  // namespace operators
+}  // namespace paddle
diff --git a/paddle/operators/sgd_op.cc b/paddle/operators/sgd_op.cc
index ad267e7f087943ff3b8326a7baf2ce3955fa51c2..1232e64c7f0132b9ea19b3d7e1ebe9531e1e25a5 100644
--- a/paddle/operators/sgd_op.cc
+++ b/paddle/operators/sgd_op.cc
@@ -23,10 +23,18 @@ class SGDOp : public framework::OperatorWithKernel {
 
  protected:
   void InferShape(const framework::InferShapeContext &ctx) const override {
-    PADDLE_ENFORCE(
-        ctx.Input<Tensor>("param")->dims() == ctx.Input<Tensor>("grad")->dims(),
-        "Two input of SGD Op's dimension must be same.");
-    ctx.Output<Tensor>("param_out")->Resize(ctx.Input<Tensor>("param")->dims());
+    PADDLE_ENFORCE_NOT_NULL(ctx.InputVar("param"),
+                            "Input(param) of SGDOp should not be null.");
+    PADDLE_ENFORCE_NOT_NULL(ctx.InputVar("grad"),
+                            "Input(grad) of SGDOp should not be null.");
+    PADDLE_ENFORCE_NOT_NULL(ctx.OutputVar("param_out"),
+                            "Output(param_out) of SGDOp should not be null.");
+
+    PADDLE_ENFORCE_EQ(ctx.Input<Tensor>("param")->dims(),
+                      ctx.Input<Tensor>("grad")->dims(),
+                      "Two input of SGD Op's dimension must be same.");
+    ctx.Output<framework::LoDTensor>("param_out")
+        ->Resize(ctx.Input<Tensor>("param")->dims());
   }
 };
 
diff --git a/paddle/operators/sigmoid_op.cc b/paddle/operators/sigmoid_op.cc
index 761c6de8d4d2150b30b97b58da95da3d5f33db63..992b19965e0ca9ce7dba1b8b3c5b7780af06eb45 100644
--- a/paddle/operators/sigmoid_op.cc
+++ b/paddle/operators/sigmoid_op.cc
@@ -23,7 +23,13 @@ class SigmoidOp : public framework::OperatorWithKernel {
 
  protected:
   void InferShape(const framework::InferShapeContext &ctx) const override {
-    ctx.Output<Tensor>("Y")->Resize(ctx.Input<Tensor>("X")->dims());
+    PADDLE_ENFORCE_NOT_NULL(ctx.InputVar("X"),
+                            "Input(X) of SigmoidOp should not be null.");
+    PADDLE_ENFORCE_NOT_NULL(ctx.OutputVar("Y"),
+                            "Output(Y) of SigmoidOp should not be null.");
+
+    ctx.Output<framework::LoDTensor>("Y")->Resize(
+        ctx.Input<Tensor>("X")->dims());
   }
 };
 
@@ -44,7 +50,7 @@ class SigmoidOpGrad : public framework::OperatorWithKernel {
 
  protected:
   void InferShape(const framework::InferShapeContext &ctx) const override {
-    ctx.Output<Tensor>(framework::GradVarName("X"))
+    ctx.Output<framework::LoDTensor>(framework::GradVarName("X"))
         ->Resize(ctx.Input<Tensor>("Y")->dims());
   }
 };
diff --git a/paddle/operators/softmax_op.cc b/paddle/operators/softmax_op.cc
index 7166b2f60be8a6088ab3a81686f7bed1b7181d97..c67eb028c882ed82ca4e6a4dd70cdea9f69cdc24 100644
--- a/paddle/operators/softmax_op.cc
+++ b/paddle/operators/softmax_op.cc
@@ -23,9 +23,15 @@ class SoftmaxOp : public framework::OperatorWithKernel {
 
  protected:
   void InferShape(const framework::InferShapeContext &ctx) const override {
+    PADDLE_ENFORCE_NOT_NULL(ctx.InputVar("X"),
+                            "Input(X) of SoftmaxOp should not be null.");
+    PADDLE_ENFORCE_NOT_NULL(ctx.OutputVar("Y"),
+                            "Output(Y) of SoftmaxOp should not be null.");
+
     PADDLE_ENFORCE(ctx.Input<Tensor>("X")->dims().size() == 2UL,
                    "The input of softmax op must be a matrix.");
-    ctx.Output<Tensor>("Y")->Resize(ctx.Input<Tensor>("X")->dims());
+    ctx.Output<framework::LoDTensor>("Y")->Resize(
+        ctx.Input<Tensor>("X")->dims());
   }
 };
 
@@ -71,7 +77,7 @@ class SoftmaxOpGrad : public framework::OperatorWithKernel {
                       ctx.Input<Tensor>(framework::GradVarName("Y"))->dims(),
                       "Input(Y) and its gradients should have a same shape.");
 
-    ctx.Output<Tensor>(framework::GradVarName("X"))
+    ctx.Output<framework::LoDTensor>(framework::GradVarName("X"))
         ->Resize(ctx.Input<Tensor>("X")->dims());
   }
 };
diff --git a/paddle/operators/squared_l2_distance_op.cc b/paddle/operators/squared_l2_distance_op.cc
index 9f51d3efa8ecba894a1023b9de2df451ca85916c..39f4305877de20d451bc35fe698a0eabf9758d57 100644
--- a/paddle/operators/squared_l2_distance_op.cc
+++ b/paddle/operators/squared_l2_distance_op.cc
@@ -23,12 +23,18 @@ class SquaredL2DistanceOp : public framework::OperatorWithKernel {
 
  protected:
   void InferShape(const framework::InferShapeContext& ctx) const override {
-    PADDLE_ENFORCE_NOT_NULL(ctx.InputVar("X"),
-                            "Input of SquaredL2DistanceOp "
-                            "must be initialized.");
-    PADDLE_ENFORCE_NOT_NULL(ctx.InputVar("Y"),
-                            "Target of SquaredL2DistanceOp "
-                            "must be initialized.");
+    PADDLE_ENFORCE_NOT_NULL(
+        ctx.InputVar("X"),
+        "Input(X) of SquaredL2DistanceOp should not be null.");
+    PADDLE_ENFORCE_NOT_NULL(
+        ctx.InputVar("Y"),
+        "Input(Y) of SquaredL2DistanceOp should not be null.");
+    PADDLE_ENFORCE_NOT_NULL(
+        ctx.OutputVar("sub_result"),
+        "Output(sub_result) of SquaredL2DistanceOp should not be null.");
+    PADDLE_ENFORCE_NOT_NULL(
+        ctx.OutputVar("Out"),
+        "Output(Out) of SquaredL2DistanceOp should not be null.");
 
     auto* x = ctx.Input<Tensor>("X");
     auto x_dims = x->dims();
@@ -48,9 +54,9 @@ class SquaredL2DistanceOp : public framework::OperatorWithKernel {
                    "First dimension of target must be equal to input "
                    "or to 1.");
 
-    ctx.Output<Tensor>("sub_result")
+    ctx.Output<framework::LoDTensor>("sub_result")
         ->Resize({x_dims[0], x->numel() / x_dims[0]});
-    ctx.Output<Tensor>("Out")->Resize({x_dims[0], 1});
+    ctx.Output<framework::LoDTensor>("Out")->Resize({x_dims[0], 1});
   }
 };
 
@@ -94,8 +100,10 @@ class SquaredL2DistanceGradOp : public framework::OperatorWithKernel {
     PADDLE_ENFORCE_EQ(out_dims[1], 1,
                       "Second dimension of output gradient "
                       "must be 1.");
-    auto* x_grad = ctx.Output<Tensor>(framework::GradVarName("X"));
-    auto* y_grad = ctx.Output<Tensor>(framework::GradVarName("Y"));
+    auto* x_grad =
+        ctx.Output<framework::LoDTensor>(framework::GradVarName("X"));
+    auto* y_grad =
+        ctx.Output<framework::LoDTensor>(framework::GradVarName("Y"));
     if (x_grad) x_grad->Resize(x_dims);
     if (y_grad) y_grad->Resize(y_dims);
   }
diff --git a/paddle/operators/sum_op.cc b/paddle/operators/sum_op.cc
index 5805826ee8a555ca6dfc1ca81feaadffea9e1012..41e05c27f9029b2664685d3979fadcfd2bf6dbce 100644
--- a/paddle/operators/sum_op.cc
+++ b/paddle/operators/sum_op.cc
@@ -22,8 +22,13 @@ class SumOp : public framework::OperatorWithKernel {
 
  protected:
   void InferShape(const framework::InferShapeContext &ctx) const override {
+    PADDLE_ENFORCE(!ctx.MultiInputVar("X").empty(),
+                   "Input(X) of SumOp should not be null.");
+    PADDLE_ENFORCE_NOT_NULL(ctx.OutputVar("Out"),
+                            "Output(Out) of SumOp should not be null.");
+
     auto ins = ctx.MultiInput<framework::Tensor>("X");
-    auto *out = ctx.Output<framework::Tensor>("Out");
+    auto *out = ctx.Output<framework::LoDTensor>("Out");
     int N = ins.size();
 
     auto in_dim = ins[0]->dims();
@@ -55,7 +60,8 @@ class SumGradOp : public framework::OperatorWithKernel {
 
  protected:
   void InferShape(const framework::InferShapeContext &ctx) const override {
-    auto outputs = ctx.MultiOutput<Tensor>(framework::GradVarName("X"));
+    auto outputs =
+        ctx.MultiOutput<framework::LoDTensor>(framework::GradVarName("X"));
     auto dims = ctx.Input<Tensor>(framework::GradVarName("Out"))->dims();
     for (auto output : outputs) {
       output->Resize(dims);
diff --git a/paddle/operators/top_k_op.cc b/paddle/operators/top_k_op.cc
index 38d2f0a09aec751734864947a2f3cfa20107e22f..169b815feffd86f9ff04c129ccc997230ce03a8c 100644
--- a/paddle/operators/top_k_op.cc
+++ b/paddle/operators/top_k_op.cc
@@ -24,7 +24,12 @@ class TopkOp : public framework::OperatorWithKernel {
  protected:
   void InferShape(const framework::InferShapeContext &ctx) const override {
     PADDLE_ENFORCE_NOT_NULL(ctx.InputVar("X"),
-                            "Input of TopkOP must be initialized.");
+                            "Input(X) of TopkOp should not be null.");
+    PADDLE_ENFORCE_NOT_NULL(ctx.OutputVar("Out"),
+                            "Output(Out) of TopkOp should not be null.");
+    PADDLE_ENFORCE_NOT_NULL(ctx.OutputVar("Indices"),
+                            "Output(Indices) of TopkOp should not be null.");
+
     auto *input = ctx.Input<framework::Tensor>("X");
     const int k = static_cast<int>(ctx.Attr<int>("k"));
 
@@ -35,8 +40,8 @@ class TopkOp : public framework::OperatorWithKernel {
 
     framework::DDim dims = input->dims();
     dims[dims.size() - 1] = k;
-    ctx.Output<Tensor>("Out")->Resize(dims);
-    ctx.Output<Tensor>("Indices")->Resize(dims);
+    ctx.Output<framework::LoDTensor>("Out")->Resize(dims);
+    ctx.Output<framework::LoDTensor>("Indices")->Resize(dims);
   }
 };
 
diff --git a/paddle/operators/uniform_random_op.cc b/paddle/operators/uniform_random_op.cc
index b8fbc9b52aecdb5c8d985b5de9bcd7cb85835b60..184bcbc29c0d26a214345506f126f9cc0d406b07 100644
--- a/paddle/operators/uniform_random_op.cc
+++ b/paddle/operators/uniform_random_op.cc
@@ -48,9 +48,13 @@ class UniformRandomOp : public framework::OperatorWithKernel {
 
  protected:
   void InferShape(const framework::InferShapeContext& ctx) const override {
+    PADDLE_ENFORCE_NOT_NULL(
+        ctx.OutputVar("Out"),
+        "Output(Out) of UniformRandomOp should not be null.");
+
     PADDLE_ENFORCE(Attr<float>("min") < Attr<float>("max"),
                    "uniform_random's min must less then max");
-    auto* tensor = ctx.Output<framework::Tensor>("Out");
+    auto* tensor = ctx.Output<framework::LoDTensor>("Out");
     auto dims = Attr<std::vector<int>>("dims");
     std::vector<int64_t> temp;
     temp.reserve(dims.size());
diff --git a/paddle/platform/CMakeLists.txt b/paddle/platform/CMakeLists.txt
index 17bdac8749e31565b119b2cb84aed199fac0f441..8b605e51c3f4ea38fc358ce054bb36fcc82063c4 100644
--- a/paddle/platform/CMakeLists.txt
+++ b/paddle/platform/CMakeLists.txt
@@ -24,3 +24,4 @@ cc_library(device_context SRCS device_context.cc DEPS memory buddy_allocator
 nv_test(device_context_test SRCS device_context_test.cc DEPS device_context gpu_info)
 
 nv_test(cudnn_helper_test SRCS cudnn_helper_test.cc DEPS dynload_cuda)
+nv_test(transform_test SRCS transform_test.cu DEPS paddle_memory place)
diff --git a/paddle/platform/details/device_ptr_cast.h b/paddle/platform/details/device_ptr_cast.h
new file mode 100644
index 0000000000000000000000000000000000000000..4015491fcdc3554029aa771ab7da1b2f3424321f
--- /dev/null
+++ b/paddle/platform/details/device_ptr_cast.h
@@ -0,0 +1,56 @@
+/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+    http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License. */
+
+#pragma once
+
+#ifndef __NVCC__
+#error device_ptr_cast must be include by .cu file
+#endif
+
+#include <thrust/device_ptr.h>
+
+namespace paddle {
+namespace platform {
+namespace details {
+template <typename T, bool is_ptr>
+struct DevicePtrCast;
+
+template <typename T>
+struct DevicePtrCast<T, true> {
+  using ELEM = typename std::remove_pointer<T>::type;
+  using RTYPE = thrust::device_ptr<ELEM>;
+
+  inline thrust::device_ptr<ELEM> operator()(ELEM* ele) const {
+    return thrust::device_pointer_cast(ele);
+  }
+};
+
+template <typename T>
+struct DevicePtrCast<T, false> {
+  using RTYPE = T;
+  inline RTYPE operator()(RTYPE it) const { return it; }
+};
+
+// Cast T to thrust::device_ptr if T is a pointer.
+// Otherwise, e.g., T is a iterator, return T itself.
+template <typename T>
+auto DevPtrCast(T t) ->
+    typename DevicePtrCast<T, std::is_pointer<T>::value>::RTYPE {
+  DevicePtrCast<T, std::is_pointer<T>::value> cast;
+  return cast(t);
+}
+
+}  // namespace details
+}  // namespace platform
+}  // namespace paddle
diff --git a/paddle/platform/enforce.h b/paddle/platform/enforce.h
index 64fcbd93b6c4d5d9b36f2636c3ef4f7327f08d25..df5f71ed760952ed042d7ffa40a4319a73fb93bf 100644
--- a/paddle/platform/enforce.h
+++ b/paddle/platform/enforce.h
@@ -25,6 +25,10 @@ limitations under the License. */
 #include "paddle/string/printf.h"
 #include "paddle/string/to_string.h"
 
+#ifdef __GNUC__
+#include <cxxabi.h>  // for __cxa_demangle
+#endif
+
 #ifndef PADDLE_ONLY_CPU
 
 #include "paddle/platform/dynload/cublas.h"
@@ -42,6 +46,19 @@ limitations under the License. */
 namespace paddle {
 namespace platform {
 
+namespace {
+#ifdef __GNUC__
+inline std::string demangle(std::string name) {
+  int status = -4;  // some arbitrary value to eliminate the compiler warning
+  std::unique_ptr<char, void (*)(void*)> res{
+      abi::__cxa_demangle(name.c_str(), NULL, NULL, &status), std::free};
+  return (status == 0) ? res.get() : name;
+}
+#else
+inline std::string demangle(std::string name) { return name; }
+#endif
+}
+
 struct EnforceNotMet : public std::exception {
   std::exception_ptr exp_;
   std::string err_str_;
@@ -61,8 +78,8 @@ struct EnforceNotMet : public std::exception {
 
       Dl_info info;
       for (int i = 0; i < size; ++i) {
-        if (dladdr(call_stack[i], &info)) {
-          auto demangled = info.dli_sname;
+        if (dladdr(call_stack[i], &info) && info.dli_sname) {
+          auto demangled = demangle(info.dli_sname);
           auto addr_offset = static_cast<char*>(call_stack[i]) -
                              static_cast<char*>(info.dli_saddr);
           sout << string::Sprintf("%-3d %*0p %s + %zd\n", i,
diff --git a/paddle/platform/transform.h b/paddle/platform/transform.h
new file mode 100644
index 0000000000000000000000000000000000000000..3ee4acd29660f201d318ce6d39baa6f3999ae274
--- /dev/null
+++ b/paddle/platform/transform.h
@@ -0,0 +1,66 @@
+/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve.
+
+   Licensed under the Apache License, Version 2.0 (the "License");
+   you may not use this file except in compliance with the License.
+   You may obtain a copy of the License at
+
+   http://www.apache.org/licenses/LICENSE-2.0
+
+   Unless required by applicable law or agreed to in writing, software
+   distributed under the License is distributed on an "AS IS" BASIS,
+   WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+   See the License for the specific language governing permissions and
+   limitations under the License. */
+
+#pragma once
+
+#include "paddle/platform/enforce.h"
+#include "paddle/platform/hostdevice.h"
+#include "paddle/platform/place.h"
+
+#include <algorithm>
+#include <type_traits>
+#ifdef __NVCC__
+#include <thrust/transform.h>
+#include "paddle/platform/details/device_ptr_cast.h"
+#endif
+
+namespace paddle {
+namespace platform {
+// Transform on host or device. It provides the same API in std library.
+template <typename Place, typename InputIter, typename OutputIter,
+          typename UnaryOperation>
+void Transform(Place place, InputIter first, InputIter last, OutputIter result,
+               UnaryOperation op) {
+  if (is_cpu_place(place)) {
+    std::transform(first, last, result, op);
+  } else {
+#ifdef __NVCC__
+    using namespace details;
+    thrust::transform(DevPtrCast(first), DevPtrCast(last), DevPtrCast(result),
+                      op);
+#else
+    PADDLE_THROW("Do not invoke `Transform<GPUPlace>` in .cc file");
+#endif
+  }
+}
+
+template <typename Place, typename InputIter1, typename InputIter2,
+          typename OutputIter, typename BinaryOperation>
+void Transform(Place place, InputIter1 first1, InputIter1 last1,
+               InputIter2 first2, OutputIter result, BinaryOperation op) {
+  if (is_cpu_place(place)) {
+    std::transform(first1, last1, first2, result, op);
+  } else {
+#ifdef __NVCC__
+    using namespace details;
+    thrust::transform(DevPtrCast(first1), DevPtrCast(last1), DevPtrCast(first2),
+                      DevPtrCast(result), op);
+#else
+    PADDLE_THROW("Do not invoke `Transform<GPUPlace>` in .cc file");
+#endif
+  }
+};
+
+}  // namespace platform
+}  // namespace paddle
diff --git a/paddle/platform/transform_test.cu b/paddle/platform/transform_test.cu
new file mode 100644
index 0000000000000000000000000000000000000000..600fed8f45077a6fee91f295aa854153c9cf9c01
--- /dev/null
+++ b/paddle/platform/transform_test.cu
@@ -0,0 +1,84 @@
+/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve.
+
+   Licensed under the Apache License, Version 2.0 (the "License");
+   you may not use this file except in compliance with the License.
+   You may obtain a copy of the License at
+
+   http://www.apache.org/licenses/LICENSE-2.0
+
+   Unless required by applicable law or agreed to in writing, software
+   distributed under the License is distributed on an "AS IS" BASIS,
+   WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+   See the License for the specific language governing permissions and
+   limitations under the License. */
+
+#include <gtest/gtest.h>
+#include "paddle/memory/memcpy.h"
+#include "paddle/memory/memory.h"
+#include "paddle/platform/transform.h"
+
+template <typename T>
+class Scale {
+ public:
+  explicit Scale(const T& scale) : scale_(scale) {}
+
+  HOSTDEVICE T operator()(const T& a) const { return a * scale_; }
+
+ private:
+  T scale_;
+};
+
+template <typename T>
+class Multiply {
+ public:
+  HOSTDEVICE T operator()(const T& a, const T& b) const { return a * b; }
+};
+
+TEST(Transform, CPUUnary) {
+  using namespace paddle::platform;
+  float buf[4] = {0.1, 0.2, 0.3, 0.4};
+  Transform(CPUPlace(), buf, buf + 4, buf, Scale<float>(10));
+  for (int i = 0; i < 4; ++i) {
+    ASSERT_NEAR(buf[i], static_cast<float>(i + 1), 1e-5);
+  }
+}
+
+TEST(Transform, GPUUnary) {
+  using namespace paddle::platform;
+  using namespace paddle::memory;
+  GPUPlace gpu0(0);
+  float cpu_buf[4] = {0.1, 0.2, 0.3, 0.4};
+  float* gpu_buf = static_cast<float*>(Alloc(gpu0, sizeof(float) * 4));
+  Copy(gpu0, gpu_buf, CPUPlace(), cpu_buf, sizeof(cpu_buf));
+  Transform(gpu0, gpu_buf, gpu_buf + 4, gpu_buf, Scale<float>(10));
+  Copy(CPUPlace(), cpu_buf, gpu0, gpu_buf, sizeof(cpu_buf));
+  Free(gpu0, gpu_buf);
+  for (int i = 0; i < 4; ++i) {
+    ASSERT_NEAR(cpu_buf[i], static_cast<float>(i + 1), 1e-5);
+  }
+}
+
+TEST(Transform, CPUBinary) {
+  using namespace paddle::platform;
+  using namespace paddle::memory;
+  int buf[4] = {1, 2, 3, 4};
+  Transform(CPUPlace(), buf, buf + 4, buf, buf, Multiply<int>());
+  for (int i = 0; i < 4; ++i) {
+    ASSERT_EQ((i + 1) * (i + 1), buf[i]);
+  }
+}
+
+TEST(Transform, GPUBinary) {
+  using namespace paddle::platform;
+  using namespace paddle::memory;
+  int buf[4] = {1, 2, 3, 4};
+  GPUPlace gpu0(0);
+  int* gpu_buf = static_cast<int*>(Alloc(gpu0, sizeof(buf)));
+  Copy(gpu0, gpu_buf, CPUPlace(), buf, sizeof(buf));
+  Transform(gpu0, gpu_buf, gpu_buf + 4, gpu_buf, gpu_buf, Multiply<int>());
+  Copy(CPUPlace(), buf, gpu0, gpu_buf, sizeof(buf));
+  Free(gpu0, gpu_buf);
+  for (int i = 0; i < 4; ++i) {
+    ASSERT_EQ((i + 1) * (i + 1), buf[i]);
+  }
+}
\ No newline at end of file
diff --git a/paddle/pybind/pybind.cc b/paddle/pybind/pybind.cc
index 13e11fe82a351ee9697f3e9c4e115a00e557d212..c7009a604f60cda11434ad33b6c7d7caee1befdd 100644
--- a/paddle/pybind/pybind.cc
+++ b/paddle/pybind/pybind.cc
@@ -19,10 +19,12 @@ limitations under the License. */
 #include "paddle/framework/backward.h"
 #include "paddle/framework/lod_tensor.h"
 #include "paddle/framework/op_registry.h"
+#include "paddle/operators/cond_op.h"
 #include "paddle/operators/net_op.h"
 #include "paddle/operators/recurrent_op.h"
 #include "paddle/platform/enforce.h"
 #include "paddle/platform/place.h"
+#include "paddle/pybind/pybind.h"
 #include "paddle/pybind/tensor_py.h"
 #include "paddle/string/to_string.h"
 #include "pybind11/numpy.h"
@@ -31,31 +33,6 @@ limitations under the License. */
 
 namespace py = pybind11;
 
-USE_OP(add);
-USE_OP(cross_entropy);
-USE_OP(sgd);
-USE_OP(mul);
-USE_OP(mean);
-USE_OP(sigmoid);
-USE_OP(softmax);
-USE_OP(rowwise_add);
-USE_OP(fill_zeros_like);
-USE_NO_KERNEL_OP(recurrent);
-USE_OP(gaussian_random);
-USE_OP(uniform_random);
-USE_OP(lookup_table);
-USE_OP(scale);
-USE_NO_KERNEL_OP(identity);
-USE_OP(minus);
-USE_OP(cos_sim);
-USE_CPU_ONLY_OP(gather);
-USE_CPU_ONLY_OP(scatter);
-USE_CPU_ONLY_OP(concat);
-USE_OP(top_k);
-USE_OP(squared_l2_distance);
-USE_OP(sum);
-USE_OP(reshape);
-
 namespace paddle {
 namespace framework {
 
@@ -121,27 +98,21 @@ PYBIND11_PLUGIN(core) {
         return self.data<float>()[offset];
       });
 
-  py::class_<LoDTensor>(m, "LoDTensor", R"DOC(LoD(Leval of Ddetails) Tensor.
-
-The tensor and LoD info should be created before creating the LoDTensor, then
-call the set_tensor and set_lod functions to set them.
-
-)DOC")
-      .def("__init__",
-           [](LoDTensor &instance,
-              const std::vector<std::vector<size_t>> &lod,
-              Tensor *t) {
+  py::class_<LoDTensor, Tensor>(m, "LoDTensor")
+      .def_buffer(
+          [](Tensor &self) -> py::buffer_info { return CastToPyBuffer(self); })
+      .def(
+          "__init__",
+          [](LoDTensor &instance, const std::vector<std::vector<size_t>> &lod) {
 #ifdef PADDLE_ONLY_CPU
-             new (&instance) LoDTensor(lod, t);
+            new (&instance) LoDTensor(lod);
 #else
              paddle::framework::LoD new_lod;
              new_lod.reserve(lod.size());
              std::copy(lod.begin(), lod.end(), std::back_inserter(new_lod));
-             new (&instance) LoDTensor(new_lod, t);
+             new (&instance) LoDTensor(new_lod);
 #endif
-           })
-      .def("set_tensor",
-           [](LoDTensor &self, Tensor *tensor) { self.set_tensor(tensor); })
+          })
       .def("set_lod",
            [](LoDTensor &self, const std::vector<std::vector<size_t>> &lod) {
 #ifdef PADDLE_ONLY_CPU
@@ -153,9 +124,6 @@ call the set_tensor and set_lod functions to set them.
              self.set_lod(new_lod);
 #endif
            })
-      .def("tensor",
-           [](LoDTensor &self) -> Tensor & { return self.tensor(); },
-           py::return_value_policy::reference)
       .def("lod", [](LoDTensor &self) -> std::vector<std::vector<size_t>> {
 #ifdef PADDLE_ONLY_CPU
         return self.lod();
@@ -184,9 +152,6 @@ All parameter, weight, gradient are variables in Paddle.
            [](Variable &var, int val) -> void { *var.GetMutable<int>() = val; })
       .def("get_int", [](const Variable &var) -> int { return var.Get<int>(); })
       .def("get_tensor",
-           [](Variable &self) -> Tensor * { return self.GetMutable<Tensor>(); },
-           py::return_value_policy::reference)
-      .def("get_lod_tensor",
            [](Variable &self) -> LoDTensor * {
              return self.GetMutable<LoDTensor>();
            },
@@ -324,6 +289,28 @@ All parameter, weight, gradient are variables in Paddle.
            [](operators::RecurrentOp &self, const operators::NetOp &net)
                -> void { self.set_stepnet(net.Clone()); });
 
+  // cond_op
+  py::class_<operators::CondOp, OperatorBase>(m, "CondOp")
+      .def_static("create",
+                  [](py::bytes protobin) -> operators::CondOp * {
+                    OpDesc desc;
+                    PADDLE_ENFORCE(desc.ParsePartialFromString(protobin),
+                                   "Cannot parse user input to OpDesc");
+                    PADDLE_ENFORCE(desc.IsInitialized(),
+                                   "User OpDesc is not initialized, reason %s",
+                                   desc.InitializationErrorString());
+                    auto cond_op = OpRegistry::CreateOp(desc);
+                    return static_cast<operators::CondOp *>(cond_op.release());
+                  })
+      .def("set_truenet",
+           [](operators::CondOp &self, const operators::NetOp &net) -> void {
+             self.set_truenet(net.Clone());
+           })
+      .def("set_falsenet",
+           [](operators::CondOp &self, const operators::NetOp &net) -> void {
+             self.set_falsenet(net.Clone());
+           });
+
   m.def("unique_integer", UniqueIntegerGenerator);
 
   m.def("is_compile_gpu", IsCompileGPU);
diff --git a/paddle/scripts/docker/build_android.sh b/paddle/scripts/docker/build_android.sh
index aabd2da5e499c8e648f2967e56c661ec37f025a1..11612ad4bed0afa8496087605afaefbd0420d5ce 100644
--- a/paddle/scripts/docker/build_android.sh
+++ b/paddle/scripts/docker/build_android.sh
@@ -2,8 +2,30 @@
 
 set -xe
 
+if [ $ANDROID_ABI == "arm64-v8a" ]; then
+  ANDROID_ARCH=arm64
+else # armeabi, armeabi-v7a
+  ANDROID_ARCH=arm
+fi
+
+ANDROID_STANDALONE_TOOLCHAIN=$ANDROID_TOOLCHAINS_DIR/$ANDROID_ARCH-android-$ANDROID_API
+
+cat <<EOF
+============================================
+Generating the standalone toolchain ...
+${ANDROID_NDK_HOME}/build/tools/make-standalone-toolchain.sh
+      --arch=$ANDROID_ARCH
+      --platform=android-$ANDROID_API
+      --install-dir=${ANDROID_STANDALONE_TOOLCHAIN}
+============================================
+EOF
+${ANDROID_NDK_HOME}/build/tools/make-standalone-toolchain.sh \
+      --arch=$ANDROID_ARCH \
+      --platform=android-$ANDROID_API \
+      --install-dir=$ANDROID_STANDALONE_TOOLCHAIN
+
 BUILD_ROOT=/paddle/build_android
-DEST_ROOT=/paddle/install
+DEST_ROOT=/paddle/install_android
 
 rm -rf $BUILD_ROOT 2>/dev/null || true
 mkdir -p $BUILD_ROOT
@@ -11,7 +33,7 @@ cd $BUILD_ROOT
 
 if [ $ANDROID_ABI == "armeabi-v7a" ]; then
   cmake -DCMAKE_SYSTEM_NAME=Android \
-        -DANDROID_STANDALONE_TOOLCHAIN=$ANDROID_ARM_STANDALONE_TOOLCHAIN \
+        -DANDROID_STANDALONE_TOOLCHAIN=$ANDROID_STANDALONE_TOOLCHAIN \
         -DANDROID_ABI=$ANDROID_ABI \
         -DANDROID_ARM_NEON=ON \
         -DANDROID_ARM_MODE=ON \
@@ -26,7 +48,7 @@ if [ $ANDROID_ABI == "armeabi-v7a" ]; then
         ..
 elif [ $ANDROID_ABI == "arm64-v8a" ]; then
   cmake -DCMAKE_SYSTEM_NAME=Android \
-        -DANDROID_STANDALONE_TOOLCHAIN=$ANDROID_ARM64_STANDALONE_TOOLCHAIN \
+        -DANDROID_STANDALONE_TOOLCHAIN=$ANDROID_STANDALONE_TOOLCHAIN \
         -DANDROID_ABI=$ANDROID_ABI \
         -DANDROID_ARM_MODE=ON \
         -DHOST_C_COMPILER=/usr/bin/gcc \
@@ -40,12 +62,12 @@ elif [ $ANDROID_ABI == "arm64-v8a" ]; then
         ..
 elif [ $ANDROID_ABI == "armeabi" ]; then
   cmake -DCMAKE_SYSTEM_NAME=Android \
-        -DANDROID_STANDALONE_TOOLCHAIN=$ANDROID_ARM_STANDALONE_TOOLCHAIN \
+        -DANDROID_STANDALONE_TOOLCHAIN=$ANDROID_STANDALONE_TOOLCHAIN \
         -DANDROID_ABI=$ANDROID_ABI \
         -DANDROID_ARM_MODE=ON \
         -DHOST_C_COMPILER=/usr/bin/gcc \
         -DHOST_CXX_COMPILER=/usr/bin/g++ \
-        -DCMAKE_INSTALL_PREFIX=/paddle/install \
+        -DCMAKE_INSTALL_PREFIX=$DEST_ROOT \
         -DCMAKE_BUILD_TYPE=Release \
         -DWITH_C_API=ON \
         -DWITH_SWIG_PY=OFF \
@@ -55,5 +77,10 @@ else
   echo "Invalid ANDROID_ABI: $ANDROID_ABI"
 fi
 
+cat <<EOF
+============================================
+Building in $BUILD_ROOT ...
+============================================
+EOF
 make -j `nproc`
 make install -j `nproc`
diff --git a/python/paddle/trainer/config_parser.py b/python/paddle/trainer/config_parser.py
index 4f68a8953446ffa0510df65c5b214d09b913cff8..a9e1d6d2e06d56f837690ec95fa8f8d41a90725f 100644
--- a/python/paddle/trainer/config_parser.py
+++ b/python/paddle/trainer/config_parser.py
@@ -2055,20 +2055,26 @@ class ConvLayerBase(LayerBase):
         if num_filters is not None:
             self.config.num_filters = num_filters
 
+        use_mkldnn = int(g_command_config_args.get("use_mkldnn", 0))
         use_gpu = int(g_command_config_args.get("use_gpu", 0))
         parallel_nn = int(g_command_config_args.get("parallel_nn", 0))
 
-        # Automatically select cudnn_type for GPU and exconv for CPU
+        # Automatically select cudnn_type for GPU, exconv for CPU
+        # and mkldnn_conv for MKLDNN
         # if set type=conv, but still reserve the way user specify
-        # exconv or cudnn_conv manually.
+        # exconv, mkldnn_conv or cudnn_conv manually.
         if self.layer_type == "cudnn_conv":
             config_assert(use_gpu, "cudnn_conv only support GPU")
 
+        if self.layer_type == "mkldnn_conv":
+            config_assert(use_mkldnn, "mkldnn_conv only support MKLDNN")
+
         if (use_gpu == 1 and self.layer_type != "exconv" and
+                self.layer_type != "mkldnn_conv" and
             (parallel_nn == 0 or self.config.device > -1)):
             self.layer_type = "cudnn_conv"
         else:
-            self.layer_type = "exconv"
+            self.layer_type = "mkldnn_conv" if use_mkldnn else "exconv"
         # need to specify layer in config
         self.config.type = self.layer_type
 
@@ -2100,6 +2106,11 @@ class ConvLayer(ConvLayerBase):
     layer_type = 'exconv'
 
 
+@config_layer('mkldnn_conv')
+class ConvLayer(ConvLayerBase):
+    layer_type = 'mkldnn_conv'
+
+
 @config_layer('cudnn_conv')
 class ConvLayer(ConvLayerBase):
     layer_type = 'cudnn_conv'
diff --git a/python/paddle/trainer_config_helpers/layers.py b/python/paddle/trainer_config_helpers/layers.py
index 4b1d80d3db924bfa2ad0e081f785d8f5dd719fce..8c7d1738ad9753eb7afb27e893f979f8bce70a0d 100644
--- a/python/paddle/trainer_config_helpers/layers.py
+++ b/python/paddle/trainer_config_helpers/layers.py
@@ -169,6 +169,7 @@ class LayerType(object):
     EXCONV_LAYER = 'exconv'
     EXCONVTRANS_LAYER = 'exconvt'
     CUDNNCONV_LAYER = 'cudnn_conv'
+    CUDNNCONVTRANS_LAYER = 'cudnn_convt'
     POOL_LAYER = 'pool'
     POOL3D_LAYER = 'pool3d'
     BATCH_NORM_LAYER = 'batch_norm'
diff --git a/python/paddle/v2/framework/op.py b/python/paddle/v2/framework/op.py
index 9e665adad2d3ad91d183c6815fbd7135ac4e8965..6cca41e43b38b8cccb65ff9b347ef226dddecd4d 100644
--- a/python/paddle/v2/framework/op.py
+++ b/python/paddle/v2/framework/op.py
@@ -97,7 +97,7 @@ class OpDescCreationMethod(object):
                     new_attr.strings.extend(user_defined_attr)
                 elif attr.type == framework_pb2.INT_PAIRS:
                     for p in user_defined_attr:
-                        pair = new_attr.pairs.add()
+                        pair = new_attr.int_pairs.add()
                         pair.first = p[0]
                         pair.second = p[1]
                 else:
@@ -215,5 +215,27 @@ class __RecurrentOp__(object):
         return core.RecurrentOp.create(proto.SerializeToString())
 
 
+class __CondOp__(object):
+    __proto__ = None
+    type = "cond"
+
+    def __init__(self):
+        # cache recurrent_op's proto
+        if self.__proto__ is None:
+            for op_proto in get_all_op_protos():
+                if op_proto.type == self.type:
+                    self.__proto__ = op_proto
+
+    def __call__(self, *args, **kwargs):
+        if self.type not in args and "type" not in kwargs:
+            kwargs["type"] = self.type
+        # create proto
+        create_method = OpDescCreationMethod(self.__proto__)
+        proto = create_method(*args, **kwargs)
+        # create condop
+        return core.CondOp.create(proto.SerializeToString())
+
+
 Operator = OperatorFactory()  # The default global factory
 RecurrentOp = __RecurrentOp__()
+CondOp = __CondOp__()
diff --git a/python/paddle/v2/framework/tests/op_test.py b/python/paddle/v2/framework/tests/op_test.py
index 4fec4c9109bf247abb2068177583acb47a8ebd97..72e6aefb5424e4f8db488f4a79cc4c0b2bfc330f 100644
--- a/python/paddle/v2/framework/tests/op_test.py
+++ b/python/paddle/v2/framework/tests/op_test.py
@@ -47,17 +47,24 @@ def set_input(scope, op, inputs, place):
         if in_name in inputs:
             if in_dup:
                 sub_in = inputs[in_name]
-                for sub_in_name, sub_in_array in sub_in:
+                for sub_in_name, sub_in_val in sub_in:
                     var = scope.find_var(sub_in_name)
                     tensor = var.get_tensor()
+                    sub_in_array = sub_in_val[0] \
+                        if isinstance(sub_in_val, tuple) else sub_in_val
                     tensor.set_dims(sub_in_array.shape)
                     tensor.set(sub_in_array, place)
+                    if isinstance(sub_in_val, tuple):
+                        tensor.set_lod(sub_in_val[1])
             else:
                 var = scope.find_var(in_name)
                 tensor = var.get_tensor()
-                arr = inputs[in_name]
-                tensor.set_dims(arr.shape)
-                tensor.set(arr, place)
+                in_val = inputs[in_name]
+                in_array = in_val[0] if isinstance(in_val, tuple) else in_val
+                tensor.set_dims(in_array.shape)
+                tensor.set(in_array, place)
+                if isinstance(in_val, tuple):
+                    tensor.set_lod(in_val[1])
 
 
 def set_output_grad(scope, op, outputs, place):
@@ -85,7 +92,7 @@ def get_numeric_gradient(scope,
                          op,
                          inputs,
                          input_to_check,
-                         output_name,
+                         output_names,
                          delta=0.005,
                          in_place=False):
 
@@ -100,8 +107,11 @@ def get_numeric_gradient(scope,
     ctx = core.DeviceContext.create(core.CPUPlace())
 
     def get_output():
-        op.run(scope, ctx)
-        return np.array(scope.find_var(output_name).get_tensor()).sum()
+        sum = 0.0
+        for output_name in output_names:
+            op.run(scope, ctx)
+            sum += np.array(scope.find_var(output_name).get_tensor()).sum()
+        return sum
 
     tensor_to_check = scope.find_var(input_to_check).get_tensor()
     tensor_size = product(tensor_to_check.get_dims())
@@ -189,14 +199,14 @@ class OpTest(unittest.TestCase):
                     self.assertTrue(
                         np.allclose(
                             actual, expect, atol=1e-05),
-                        "output name: " + out_name + "has diff")
+                        "output name: " + out_name + " has diff")
             else:
                 actual = np.array(self.scope.find_var(out_name).get_tensor())
                 expect = self.outputs[out_name]
                 self.assertTrue(
                     np.allclose(
                         actual, expect, atol=1e-05),
-                    "output name: " + out_name + "has diff")
+                    "output name: " + out_name + " has diff")
 
     def check_output(self):
         places = [core.CPUPlace()]
@@ -225,7 +235,7 @@ class OpTest(unittest.TestCase):
 
     def check_grad(self,
                    inputs_to_check,
-                   output_name,
+                   output_names,
                    no_grad_set=None,
                    in_place=False,
                    max_relative_error=0.005):
@@ -237,13 +247,16 @@ class OpTest(unittest.TestCase):
         if no_grad_set is None:
             no_grad_set = set()
 
+        if not type(output_names) is list:
+            output_names = [output_names]
+
         numeric_grads = [
             get_numeric_gradient(
                 self.scope,
                 self.op,
                 self.inputs,
                 input_to_check,
-                output_name,
+                output_names,
                 in_place=in_place) for input_to_check in inputs_to_check
         ]
         grad_names = [
diff --git a/python/paddle/v2/framework/tests/test_accuracy_op.py b/python/paddle/v2/framework/tests/test_accuracy_op.py
new file mode 100644
index 0000000000000000000000000000000000000000..43d60eb90d5edbd6944a11f7555f0291720dd2be
--- /dev/null
+++ b/python/paddle/v2/framework/tests/test_accuracy_op.py
@@ -0,0 +1,25 @@
+import unittest
+import numpy as np
+from op_test import OpTest
+
+
+class TestAccuracyOp(OpTest):
+    def setUp(self):
+        self.op_type = "accuracy"
+        infer = np.random.randint(0, 2, (32, 1)).astype("int")
+        label = np.random.randint(0, 2, (32, )).astype("int")
+        self.inputs = {'Inference': infer, "Label": label}
+        num_correct = 0
+        for rowid in xrange(32):
+            for ele in infer[rowid]:
+                if ele == label[rowid]:
+                    num_correct += 1
+                    break
+        self.outputs = {'Accuracy': [num_correct / 32.0]}
+
+    def test_check_output(self):
+        self.check_output()
+
+
+if __name__ == '__main__':
+    unittest.main()
diff --git a/python/paddle/v2/framework/tests/test_add_two_op.py b/python/paddle/v2/framework/tests/test_add_op.py
similarity index 100%
rename from python/paddle/v2/framework/tests/test_add_two_op.py
rename to python/paddle/v2/framework/tests/test_add_op.py
diff --git a/python/paddle/v2/framework/tests/test_cond_op.py b/python/paddle/v2/framework/tests/test_cond_op.py
new file mode 100644
index 0000000000000000000000000000000000000000..37177ae0b2482517c4183969c8ef0670f2b3de89
--- /dev/null
+++ b/python/paddle/v2/framework/tests/test_cond_op.py
@@ -0,0 +1,116 @@
+import logging
+import paddle.v2.framework.core as core
+import unittest
+import numpy as np
+from paddle.v2.framework.op import Operator, CondOp
+
+
+class PySimpleCond(object):
+    '''
+    A simple implementation of dynamic if-else based on numpy
+    '''
+
+    def __init__(self):
+        array = [1] * 10
+        for i in range(1, 10, 2):
+            array[i] = 0
+        self.cond = np.array(array)
+        self.x = np.ones(shape=(10, 1))
+
+    def forward(self):
+        self.index_t = np.where(self.cond == 1)
+        self.index_f = np.where(self.cond == 0)
+        y_t = self.x[self.index_t]
+        y_f = self.x[self.index_f]
+        y_t = y_t * 2.
+        y_f = y_f * (-2.)
+        output = np.zeros(shape=(10, 1))
+        output[self.index_t] = y_t
+        output[self.index_f] = y_f
+        return output
+
+
+class PySimpleCondTest(unittest.TestCase):
+    def setUp(self):
+        self.condnn = PySimpleCond()
+
+    def test_forward(self):
+        output = self.condnn.forward()
+
+
+def create_tensor(scope, name, shape, np_data):
+    tensor = scope.new_var(name).get_tensor()
+    tensor.set_dims(shape)
+    tensor.set(np_data, core.CPUPlace())
+    return tensor
+
+
+class TestCondOp(unittest.TestCase):
+    '''
+    Test CondOp
+
+    equation:
+        cond = [True, False, True, False, ...]
+        y[index_t] = x[index_t] * 2.
+        y[index_f] = x[index_f] * -2.
+    outputs:
+        y
+    '''
+
+    def setUp(self):
+        self.py_cond = PySimpleCond()
+
+    def forward(self):
+        self.scope = core.Scope()
+        self.create_global_variables()
+        self.create_cond_op()
+        self.create_sub_net()
+        ctx = core.DeviceContext.create(core.CPUPlace())
+        self.condop.infer_shape(self.scope)
+        self.condop.run(self.scope, ctx)
+        return np.array(self.scope.find_var("Out").get_tensor())
+
+    def create_global_variables(self):
+        x_np_data = self.py_cond.x
+        create_tensor(self.scope, "X", [10, 1], x_np_data)
+        cond_np_data = self.py_cond.cond.astype("int32")
+        create_tensor(self.scope, "cond", [10, 1], cond_np_data)
+        self.scope.new_var("SubScopes")
+        self.scope.new_var("IndexTensors")
+        self.scope.new_var("Out")
+
+    def create_cond_op(self):
+        self.condop = CondOp(
+            Cond="cond",
+            Xs=["X"],
+            Outs=["Out"],
+            SubScopes="SubScopes",
+            IndexTensors="IndexTensors")
+
+    def create_sub_net(self):
+        truenet = core.Net.create()
+        scale_op_t = Operator("scale", X='X', Out='Out', scale=2.)
+        truenet.append_op(scale_op_t)
+        truenet.complete_add_op(True)
+        self.condop.set_truenet(truenet)
+
+        falsenet = core.Net.create()
+        scale_op_t = Operator("scale", X='X', Out='Out', scale=-2.)
+        falsenet.append_op(scale_op_t)
+        falsenet.complete_add_op(True)
+        self.condop.set_falsenet(falsenet)
+
+    def test_forward(self):
+        print 'test cond op forward'
+        pd_output = self.forward()
+        py_output = self.py_cond.forward()
+        print 'pd_output', pd_output
+        print
+        print 'py_output', py_output
+        self.assertEqual(pd_output.shape, py_output.shape)
+        print 'test passed'
+        return 0
+
+
+if __name__ == "__main__":
+    unittest.main()
diff --git a/python/paddle/v2/framework/tests/test_cos_sim_op.py b/python/paddle/v2/framework/tests/test_cos_sim_op.py
index 797cbd8cc5cf7f73d58ca713d02667731d5c8a0e..d314ce391ea2f10a8bd77c24e84fa3e1eebb6c73 100644
--- a/python/paddle/v2/framework/tests/test_cos_sim_op.py
+++ b/python/paddle/v2/framework/tests/test_cos_sim_op.py
@@ -7,8 +7,8 @@ class TestCosSimOp(OpTest):
     def setUp(self):
         self.op_type = "cos_sim"
         self.inputs = {
-            'X': np.random.random((10, 5)).astype("float32"),
-            'Y': np.random.random((10, 5)).astype("float32")
+            'X': np.random.random((6, 5)).astype("float32"),
+            'Y': np.random.random((6, 5)).astype("float32")
         }
         expect_x_norm = np.linalg.norm(self.inputs['X'], axis=1)
         expect_y_norm = np.linalg.norm(self.inputs['Y'], axis=1)
@@ -28,12 +28,66 @@ class TestCosSimOp(OpTest):
 
     def test_check_grad_ingore_x(self):
         self.check_grad(
-            ['Y'], 'Out', max_relative_error=0.05, no_grad_set=set('X'))
+            ['Y'], 'Out', max_relative_error=0.05, no_grad_set=set("X"))
 
-    def test_check_grad_ignore_y(self):
+    def test_check_grad_ingore_y(self):
         self.check_grad(
             ['X'], 'Out', max_relative_error=0.05, no_grad_set=set('Y'))
 
 
-if __name__ == "__main__":
+class TestCosSimOp2(TestCosSimOp):
+    def setUp(self):
+        self.op_type = "cos_sim"
+        self.inputs = {
+            'X': np.random.random((6, 5)).astype("float32"),
+            'Y': np.random.random((1, 5)).astype("float32")
+        }
+        expect_x_norm = np.linalg.norm(self.inputs['X'], axis=1)
+        expect_y_norm = np.linalg.norm(self.inputs['Y'], axis=1)
+        expect_out = (self.inputs['X'] * self.inputs['Y']).sum(axis=1) / \
+            expect_x_norm / expect_y_norm
+        self.outputs = {
+            'XNorm': np.expand_dims(expect_x_norm, 1),
+            'YNorm': np.expand_dims(expect_y_norm, 1),
+            'Out': np.expand_dims(expect_out, 1)
+        }
+
+
+class TestCosSimOp3(TestCosSimOp):
+    def setUp(self):
+        self.op_type = "cos_sim"
+        self.inputs = {
+            'X': np.random.random((6, 5, 2)).astype("float32"),
+            'Y': np.random.random((6, 5, 2)).astype("float32")
+        }
+        expect_x_norm = np.linalg.norm(self.inputs['X'], axis=(1, 2))
+        expect_y_norm = np.linalg.norm(self.inputs['Y'], axis=(1, 2))
+        expect_out = (self.inputs['X'] * self.inputs['Y']).sum(axis=(1, 2)) / \
+            expect_x_norm / expect_y_norm
+        self.outputs = {
+            'XNorm': np.expand_dims(expect_x_norm, 1),
+            'YNorm': np.expand_dims(expect_y_norm, 1),
+            'Out': np.expand_dims(expect_out, 1)
+        }
+
+
+class TestCosSimOp4(TestCosSimOp):
+    def setUp(self):
+        self.op_type = "cos_sim"
+        self.inputs = {
+            'X': np.random.random((6, 5, 2)).astype("float32"),
+            'Y': np.random.random((1, 5, 2)).astype("float32")
+        }
+        expect_x_norm = np.linalg.norm(self.inputs['X'], axis=(1, 2))
+        expect_y_norm = np.linalg.norm(self.inputs['Y'], axis=(1, 2))
+        expect_out = (self.inputs['X'] * self.inputs['Y']).sum(axis=(1, 2)) / \
+            expect_x_norm / expect_y_norm
+        self.outputs = {
+            'XNorm': np.expand_dims(expect_x_norm, 1),
+            'YNorm': np.expand_dims(expect_y_norm, 1),
+            'Out': np.expand_dims(expect_out, 1)
+        }
+
+
+if __name__ == '__main__':
     unittest.main()
diff --git a/python/paddle/v2/framework/tests/test_cross_entropy_op.py b/python/paddle/v2/framework/tests/test_cross_entropy_op.py
index b845bbc68095125a84b8f7f99c6ada65320510ec..a630dea7f54b03788b103a886e03efb2b5bb2628 100644
--- a/python/paddle/v2/framework/tests/test_cross_entropy_op.py
+++ b/python/paddle/v2/framework/tests/test_cross_entropy_op.py
@@ -3,25 +3,27 @@ import numpy
 from op_test import OpTest
 
 
-class TestCrossEntropy(OpTest):
+class TestOnehotCrossEntropyOp(OpTest):
     def setUp(self):
         self.op_type = "cross_entropy"
         batch_size = 30
         class_num = 10
+
         X = numpy.random.uniform(0.1, 1.0,
                                  [batch_size, class_num]).astype("float32")
-        label = (class_num / 2) * numpy.ones(batch_size).astype("int32")
-        self.inputs = {'X': X, 'Label': label}
-        Y = []
-        for i in range(0, batch_size):
-            Y.append(-numpy.log(X[i][label[i]]))
-        self.outputs = {'Y': numpy.array(Y).astype("float32")}
+        labels = numpy.random.randint(0, class_num, batch_size, dtype="int32")
+
+        cross_entropy = numpy.asmatrix(
+            [[-numpy.log(X[i][labels[i]])] for i in range(X.shape[0])],
+            dtype="float32")
+        self.inputs = {"X": X, "Label": labels}
+        self.outputs = {"Y": cross_entropy}
 
     def test_check_output(self):
         self.check_output()
 
     def test_check_grad(self):
-        self.check_grad(['X'], 'Y')
+        self.check_grad(["X"], "Y")
 
 
 class TestCrossEntropySoftLabel(OpTest):
@@ -35,8 +37,9 @@ class TestCrossEntropySoftLabel(OpTest):
                                      [batch_size, class_num]).astype("float32")
         label /= label.sum(axis=1, keepdims=True)
         self.inputs = {'X': X, 'Label': label}
-        Y = (-label * numpy.log(X)).sum(axis=1)
-        self.outputs = {'Y': numpy.array(Y).astype("float32")}
+        cross_entropy = (-label * numpy.log(X)).sum(
+            axis=1, keepdims=True).astype("float32")
+        self.outputs = {'Y': cross_entropy}
 
     def test_check_output(self):
         self.check_output()
diff --git a/python/paddle/v2/framework/tests/test_elementwise_mul_op.py b/python/paddle/v2/framework/tests/test_elementwise_mul_op.py
new file mode 100644
index 0000000000000000000000000000000000000000..e268cfddb26721a35ddd2d2cc18f526ff7b2f6d9
--- /dev/null
+++ b/python/paddle/v2/framework/tests/test_elementwise_mul_op.py
@@ -0,0 +1,157 @@
+import unittest
+import numpy as np
+from op_test import OpTest
+
+
+class TestElementwiseMulOp_Matrix(OpTest):
+    def setUp(self):
+        self.op_type = "elementwise_mul"
+        """ Warning
+        CPU gradient check error!
+        'X': np.random.random((32,84)).astype("float32"),
+        'Y': np.random.random((32,84)).astype("float32")
+        """
+        self.inputs = {
+            'X': np.random.uniform(0.1, 1, [13, 17]).astype("float32"),
+            'Y': np.random.uniform(0.1, 1, [13, 17]).astype("float32")
+        }
+        self.outputs = {'Out': np.multiply(self.inputs['X'], self.inputs['Y'])}
+
+    def test_check_output(self):
+        self.check_output()
+
+    def test_check_grad_normal(self):
+        self.check_grad(['X', 'Y'], 'Out', max_relative_error=0.1)
+
+    def test_check_grad_ingore_x(self):
+        self.check_grad(
+            ['Y'], 'Out', max_relative_error=0.1, no_grad_set=set("X"))
+
+    def test_check_grad_ingore_y(self):
+        self.check_grad(
+            ['X'], 'Out', max_relative_error=0.1, no_grad_set=set('Y'))
+
+
+class TestElementwiseMulOp_Vector(OpTest):
+    def setUp(self):
+        self.op_type = "elementwise_mul"
+        self.inputs = {
+            'X': np.random.random((32, )).astype("float32"),
+            'Y': np.random.random((32, )).astype("float32")
+        }
+        self.outputs = {'Out': np.multiply(self.inputs['X'], self.inputs['Y'])}
+
+    def test_check_output(self):
+        self.check_output()
+
+    def test_check_grad_normal(self):
+        self.check_grad(['X', 'Y'], 'Out', max_relative_error=0.1)
+
+    def test_check_grad_ingore_x(self):
+        self.check_grad(
+            ['Y'], 'Out', max_relative_error=0.1, no_grad_set=set("X"))
+
+    def test_check_grad_ingore_y(self):
+        self.check_grad(
+            ['X'], 'Out', max_relative_error=0.1, no_grad_set=set('Y'))
+
+
+class TestElementwiseMulOp_broadcast_0(OpTest):
+    def setUp(self):
+        self.op_type = "elementwise_mul"
+        self.inputs = {
+            'X': np.random.rand(2, 3, 4).astype(np.float32),
+            'Y': np.random.rand(2).astype(np.float32)
+        }
+
+        self.attrs = {'axis': 0}
+        self.outputs = {
+            'Out': self.inputs['X'] * self.inputs['Y'].reshape(2, 1, 1)
+        }
+
+    def test_check_output(self):
+        self.check_output()
+
+    def test_check_grad_normal(self):
+        self.check_grad(['X', 'Y'], 'Out', max_relative_error=0.1)
+
+    def test_check_grad_ingore_x(self):
+        self.check_grad(
+            ['Y'], 'Out', max_relative_error=0.1, no_grad_set=set("X"))
+
+    def test_check_grad_ingore_y(self):
+        self.check_grad(
+            ['X'], 'Out', max_relative_error=0.1, no_grad_set=set('Y'))
+
+
+class TestElementwiseMulOp_broadcast_1(OpTest):
+    def setUp(self):
+        self.op_type = "elementwise_mul"
+        self.inputs = {
+            'X': np.random.rand(2, 3, 4).astype(np.float32),
+            'Y': np.random.rand(3).astype(np.float32)
+        }
+
+        self.attrs = {'axis': 1}
+        self.outputs = {
+            'Out': self.inputs['X'] * self.inputs['Y'].reshape(1, 3, 1)
+        }
+
+    def test_check_output(self):
+        self.check_output()
+
+    def test_check_grad_normal(self):
+        self.check_grad(['X', 'Y'], 'Out', max_relative_error=0.1)
+
+    def test_check_grad_ingore_x(self):
+        self.check_grad(
+            ['Y'], 'Out', max_relative_error=0.1, no_grad_set=set("X"))
+
+    def test_check_grad_ingore_y(self):
+        self.check_grad(
+            ['X'], 'Out', max_relative_error=0.1, no_grad_set=set('Y'))
+
+
+class TestElementwiseMulOp_broadcast_2(OpTest):
+    def setUp(self):
+        self.op_type = "elementwise_mul"
+        self.inputs = {
+            'X': np.random.rand(2, 3, 4).astype(np.float32),
+            'Y': np.random.rand(4).astype(np.float32)
+        }
+
+        self.outputs = {
+            'Out': self.inputs['X'] * self.inputs['Y'].reshape(1, 1, 4)
+        }
+
+    def test_check_output(self):
+        self.check_output()
+
+    def test_check_grad_normal(self):
+        self.check_grad(['X', 'Y'], 'Out', max_relative_error=0.1)
+
+    def test_check_grad_ingore_x(self):
+        self.check_grad(
+            ['Y'], 'Out', max_relative_error=0.1, no_grad_set=set("X"))
+
+    def test_check_grad_ingore_y(self):
+        self.check_grad(
+            ['X'], 'Out', max_relative_error=0.1, no_grad_set=set('Y'))
+
+
+class TestElementwiseMulOp_broadcast_3(OpTest):
+    def setUp(self):
+        self.op_type = "elementwise_mul"
+        self.inputs = {
+            'X': np.random.rand(2, 3, 4, 5).astype(np.float32),
+            'Y': np.random.rand(3, 4).astype(np.float32)
+        }
+
+        self.attrs = {'axis': 1}
+        self.outputs = {
+            'Out': self.inputs['X'] * self.inputs['Y'].reshape(1, 3, 4, 1)
+        }
+
+
+if __name__ == '__main__':
+    unittest.main()
diff --git a/python/paddle/v2/framework/tests/test_gaussian_random_op.py b/python/paddle/v2/framework/tests/test_gaussian_random_op.py
index 1f9e4db783c9907a22db72c8a6ff06c7ca0735da..1888ee28f92c66496ce756d8a4a33d3e9ba57d7b 100644
--- a/python/paddle/v2/framework/tests/test_gaussian_random_op.py
+++ b/python/paddle/v2/framework/tests/test_gaussian_random_op.py
@@ -4,7 +4,7 @@ from paddle.v2.framework.op import Operator
 import numpy
 
 
-class GaussianRandomTest(unittest.TestCase):
+class TestGaussianRandomOp(unittest.TestCase):
     def test_cpu(self):
         self.gaussian_random_test(place=core.CPUPlace())
 
diff --git a/python/paddle/v2/framework/tests/test_gradient_checker.py b/python/paddle/v2/framework/tests/test_gradient_checker.py
index abeb01cb34158a43b5dcce5e39efc0e21e9fe638..85117bf9600975ea5d61dfb5b34335792bf6d8b2 100644
--- a/python/paddle/v2/framework/tests/test_gradient_checker.py
+++ b/python/paddle/v2/framework/tests/test_gradient_checker.py
@@ -12,7 +12,8 @@ class GetNumericGradientTest(unittest.TestCase):
         z = x + y
         scope = core.Scope()
         add_op = create_op(scope, "add", {'X': x, 'Y': y}, {'Out': z}, dict())
-        arr = get_numeric_gradient(scope, add_op, {'X': x, 'Y': y}, 'X', 'Out')
+        arr = get_numeric_gradient(scope, add_op, {'X': x,
+                                                   'Y': y}, 'X', ['Out'])
         self.assertAlmostEqual(arr.mean(), 1.0, delta=1e-4)
 
     def test_softmax_op(self):
diff --git a/python/paddle/v2/framework/tests/test_identity_op.py b/python/paddle/v2/framework/tests/test_identity_op.py
new file mode 100644
index 0000000000000000000000000000000000000000..2e95e7c786e3ff99a04b28218ec5b5decf531360
--- /dev/null
+++ b/python/paddle/v2/framework/tests/test_identity_op.py
@@ -0,0 +1,20 @@
+import unittest
+import numpy as np
+from op_test import OpTest
+
+
+class TestIdentityOp(OpTest):
+    def setUp(self):
+        self.op_type = "identity"
+        self.inputs = {'X': np.random.random((10, 10)).astype("float32")}
+        self.outputs = {'Out': self.inputs['X']}
+
+    def test_check_output(self):
+        self.check_output()
+
+    def test_check_grad(self):
+        self.check_grad(['X'], 'Out')
+
+
+if __name__ == "__main__":
+    unittest.main()
diff --git a/python/paddle/v2/framework/tests/test_lookup_table.py b/python/paddle/v2/framework/tests/test_lookup_table_op.py
similarity index 100%
rename from python/paddle/v2/framework/tests/test_lookup_table.py
rename to python/paddle/v2/framework/tests/test_lookup_table_op.py
diff --git a/python/paddle/v2/framework/tests/test_minus_op.py b/python/paddle/v2/framework/tests/test_minus_op.py
index dea797a1fea34265d0a32e097f413f421abf2521..c56d7cb548706880dd482bad750f2989c0e9a710 100644
--- a/python/paddle/v2/framework/tests/test_minus_op.py
+++ b/python/paddle/v2/framework/tests/test_minus_op.py
@@ -3,7 +3,7 @@ import numpy as np
 from op_test import OpTest
 
 
-class MinusOpTest(OpTest):
+class TestMinusOp(OpTest):
     def setUp(self):
         self.op_type = "minus"
         self.inputs = {
diff --git a/python/paddle/v2/framework/tests/test_pad_op.py b/python/paddle/v2/framework/tests/test_pad_op.py
new file mode 100644
index 0000000000000000000000000000000000000000..9052e63b5683801da7c73be4de23013c949add98
--- /dev/null
+++ b/python/paddle/v2/framework/tests/test_pad_op.py
@@ -0,0 +1,55 @@
+import unittest
+import numpy as np
+from op_test import OpTest
+
+
+class TestPadOp(OpTest):
+    def setUp(self):
+        self.initTestCase()
+        self.op_type = "pad"
+        self.inputs = {'X': np.random.random(self.shape).astype("float32"), }
+        self.attrs = {}
+        self.attrs['paddings'] = np.array(self.paddings).flatten()
+        self.attrs['pad_value'] = self.pad_value
+        self.outputs = {
+            'Out': np.pad(self.inputs['X'],
+                          self.paddings,
+                          mode='constant',
+                          constant_values=self.pad_value)
+        }
+
+    def test_check_output(self):
+        self.check_output()
+
+    def test_check_grad_normal(self):
+        self.check_grad(['X'], 'Out', max_relative_error=0.006)
+
+    def initTestCase(self):
+        self.shape = (16, 16)
+        self.paddings = [(0, 1), (2, 3)]
+        self.pad_value = 0
+
+
+class TestCase1(TestPadOp):
+    def initTestCase(self):
+        self.shape = (2, 3, 4, 4)
+        self.paddings = [(0, 1), (2, 3), (2, 1), (1, 1)]
+        self.pad_value = 0.5
+
+
+class TestCase2(TestPadOp):
+    def initTestCase(self):
+        self.shape = (2, 2, 2)
+        self.paddings = [(0, 0), (0, 0), (1, 2)]
+        self.pad_value = 1
+
+
+class TestCase3(TestPadOp):
+    def initTestCase(self):
+        self.shape = (8)
+        self.paddings = [(0, 1)]
+        self.pad_value = 0.9
+
+
+if __name__ == '__main__':
+    unittest.main()
diff --git a/python/paddle/v2/framework/tests/test_scale_and_identity_op.py b/python/paddle/v2/framework/tests/test_scale_op.py
similarity index 56%
rename from python/paddle/v2/framework/tests/test_scale_and_identity_op.py
rename to python/paddle/v2/framework/tests/test_scale_op.py
index 05d76d428299c8176d1a6adf6da15a203fa7502a..2ea1e185470280730ae8c8c0ea9568bbeb43eaf5 100644
--- a/python/paddle/v2/framework/tests/test_scale_and_identity_op.py
+++ b/python/paddle/v2/framework/tests/test_scale_op.py
@@ -3,20 +3,7 @@ import numpy as np
 from op_test import OpTest
 
 
-class IdentityTest(OpTest):
-    def setUp(self):
-        self.op_type = "identity"
-        self.inputs = {'X': np.random.random((10, 10)).astype("float32")}
-        self.outputs = {'Out': self.inputs['X']}
-
-    def test_check_output(self):
-        self.check_output()
-
-    def test_check_grad(self):
-        self.check_grad(['X'], 'Out')
-
-
-class ScaleTest(OpTest):
+class TestScaleOp(OpTest):
     def setUp(self):
         self.op_type = "scale"
         self.inputs = {'X': np.random.random((10, 10)).astype("float32")}
diff --git a/python/paddle/v2/framework/tests/test_seq_pool.py b/python/paddle/v2/framework/tests/test_seq_pool.py
new file mode 100644
index 0000000000000000000000000000000000000000..cf864936af6361da1f16df3cfb759b468214b970
--- /dev/null
+++ b/python/paddle/v2/framework/tests/test_seq_pool.py
@@ -0,0 +1,51 @@
+import unittest
+import numpy as np
+from op_test import OpTest
+
+
+class TestSeqAvgPool1D(OpTest):
+    def setUp(self):
+        self.op_type = 'sequence_avg_pool'
+        # one level, batch size is 4
+        x = np.random.uniform(0.1, 1, [11, 23]).astype('float32')
+        lod = [[0, 4, 5, 8, 11]]
+
+        out = np.zeros((4, 23)).astype('float32')
+        for i in range(4):
+            sub_x = x[lod[0][i]:lod[0][i + 1], :]
+            out[i] = sub_x.mean(axis=0)
+
+        self.inputs = {'X': (x, lod)}
+        self.outputs = {'Out': out}
+
+    def test_check_output(self):
+        self.check_output()
+
+    def test_check_grad(self):
+        self.check_grad(["X"], "Out")
+
+
+class TestSeqAvgPool2D(OpTest):
+    def setUp(self):
+        self.op_type = 'sequence_avg_pool'
+        # one level, batch size is 4
+        x = np.random.uniform(0.1, 1, [13, 3, 17]).astype('float32')
+        lod = [[0, 4, 5, 8, 13]]
+
+        out = np.zeros((4, 3, 17)).astype('float32')
+        for i in range(4):
+            sub_x = np.reshape(x[lod[0][i]:lod[0][i + 1], :], (-1, 3 * 17))
+            out[i] = np.reshape(sub_x.mean(axis=0), (3, 17))
+
+        self.inputs = {'X': (x, lod)}
+        self.outputs = {'Out': out}
+
+    def test_check_output(self):
+        self.check_output()
+
+    def test_check_grad(self):
+        self.check_grad(["X"], "Out")
+
+
+if __name__ == '__main__':
+    unittest.main()
diff --git a/python/paddle/v2/framework/tests/test_sgd_op.py b/python/paddle/v2/framework/tests/test_sgd_op.py
index 557cf15ace63e336462c7dcdbbc10f30aeedc6f4..64e54d1500c1bc134cc1efe33d41a16dbc08f2d4 100644
--- a/python/paddle/v2/framework/tests/test_sgd_op.py
+++ b/python/paddle/v2/framework/tests/test_sgd_op.py
@@ -3,7 +3,7 @@ import numpy as np
 from op_test import OpTest
 
 
-class TestSGD(OpTest):
+class TestSGDOp(OpTest):
     def setUp(self):
         self.op_type = "sgd"
         w = np.random.random((102, 105)).astype("float32")
diff --git a/python/paddle/v2/framework/tests/test_sigmoid_op.py b/python/paddle/v2/framework/tests/test_sigmoid_op.py
index 2316e49eff7bb1cdb53acb3889a6ef05060b59f3..d65d887db4af58c40e4e78fdbfd8e8ee668b7ee3 100644
--- a/python/paddle/v2/framework/tests/test_sigmoid_op.py
+++ b/python/paddle/v2/framework/tests/test_sigmoid_op.py
@@ -3,7 +3,7 @@ import numpy as np
 from op_test import OpTest
 
 
-class TestSigmoid(OpTest):
+class TestSigmoidOp(OpTest):
     def setUp(self):
         self.op_type = "sigmoid"
         self.inputs = {
diff --git a/python/paddle/v2/framework/tests/test_tensor.py b/python/paddle/v2/framework/tests/test_tensor.py
index f26ed4964c521be1cd839b39d7244f96c653cb1a..8cd93b35d7d1cb7d3b4a19e0e402ef576f1c0982 100644
--- a/python/paddle/v2/framework/tests/test_tensor.py
+++ b/python/paddle/v2/framework/tests/test_tensor.py
@@ -44,79 +44,66 @@ class TestTensor(unittest.TestCase):
         self.assertAlmostEqual(2.0, tensor_array_2[19, 11])
 
     def test_int_lod_tensor(self):
-        places = [core.CPUPlace(), core.GPUPlace(0)]
-        for place in places:
-            scope = core.Scope()
-            var = scope.new_var("test_tensor")
-            var_lod = scope.new_var("test_lod_tensor")
-
-            tensor = var.get_tensor()
-            lod_tensor = var_lod.get_lod_tensor()
-
-            tensor.set_dims([4, 4, 6])
-            tensor.alloc_int(place)
-            array = numpy.array(tensor)
-            array[0, 0, 0] = 3
-            array[3, 3, 5] = 10
-            tensor.set(array, place)
+        place = core.CPUPlace()
+        scope = core.Scope()
+        var_lod = scope.new_var("test_lod_tensor")
+        lod_tensor = var_lod.get_tensor()
 
-            lod_tensor.set_tensor(tensor)
-            lod_tensor.set_lod([[0, 2, 4]])
+        lod_tensor.set_dims([4, 4, 6])
+        lod_tensor.alloc_int(place)
+        array = numpy.array(lod_tensor)
+        array[0, 0, 0] = 3
+        array[3, 3, 5] = 10
+        lod_tensor.set(array, place)
+        lod_tensor.set_lod([[0, 2, 4]])
 
-            lod_v = numpy.array(lod_tensor.tensor())
-            self.assertTrue(numpy.alltrue(array == lod_v))
+        lod_v = numpy.array(lod_tensor)
+        self.assertTrue(numpy.alltrue(array == lod_v))
 
-            lod = lod_tensor.lod()
-            self.assertEqual(0, lod[0][0])
-            self.assertEqual(2, lod[0][1])
-            self.assertEqual(4, lod[0][2])
+        lod = lod_tensor.lod()
+        self.assertEqual(0, lod[0][0])
+        self.assertEqual(2, lod[0][1])
+        self.assertEqual(4, lod[0][2])
 
     def test_float_lod_tensor(self):
-        places = [core.CPUPlace(), core.GPUPlace(0)]
-        for place in places:
-            scope = core.Scope()
-            var = scope.new_var("test_tensor")
-            var_lod = scope.new_var("test_lod_tensor")
-
-            tensor = var.get_tensor()
-            lod_tensor = var_lod.get_lod_tensor()
-
-            tensor.set_dims([5, 2, 3, 4])
-            tensor.alloc_float(place)
+        place = core.CPUPlace()
+        scope = core.Scope()
+        var_lod = scope.new_var("test_lod_tensor")
 
-            tensor_array = numpy.array(tensor)
-            self.assertEqual((5, 2, 3, 4), tensor_array.shape)
-            tensor_array[0, 0, 0, 0] = 1.0
-            tensor_array[0, 0, 0, 1] = 2.0
-            tensor.set(tensor_array, place)
+        lod_tensor = var_lod.get_tensor()
+        lod_tensor.set_dims([5, 2, 3, 4])
+        lod_tensor.alloc_float(place)
 
-            lod_tensor.set_tensor(tensor)
+        tensor_array = numpy.array(lod_tensor)
+        self.assertEqual((5, 2, 3, 4), tensor_array.shape)
+        tensor_array[0, 0, 0, 0] = 1.0
+        tensor_array[0, 0, 0, 1] = 2.0
+        lod_tensor.set(tensor_array, place)
 
-            lod_v = numpy.array(lod_tensor.tensor())
-            self.assertAlmostEqual(1.0, lod_v[0, 0, 0, 0])
-            self.assertAlmostEqual(2.0, lod_v[0, 0, 0, 1])
-            self.assertEqual(len(lod_tensor.lod()), 0)
+        lod_v = numpy.array(lod_tensor)
+        self.assertAlmostEqual(1.0, lod_v[0, 0, 0, 0])
+        self.assertAlmostEqual(2.0, lod_v[0, 0, 0, 1])
+        self.assertEqual(len(lod_tensor.lod()), 0)
 
-            lod_py = [[0, 2, 5], [0, 2, 4, 5]]
-            lod_tensor.set_lod(lod_py)
-            lod = lod_tensor.lod()
-            self.assertListEqual(lod_py, lod)
+        lod_py = [[0, 2, 5], [0, 2, 4, 5]]
+        lod_tensor.set_lod(lod_py)
+        lod = lod_tensor.lod()
+        self.assertListEqual(lod_py, lod)
 
     def test_lod_tensor_init(self):
         scope = core.Scope()
-        var = scope.new_var("test_tensor")
         place = core.CPUPlace()
-        tensor = var.get_tensor()
-        tensor.set_dims([5, 2, 3, 4])
-        tensor.alloc_float(place)
-        tensor_array = numpy.array(tensor)
+        lod_py = [[0, 2, 5], [0, 2, 4, 5]]
+        lod_tensor = core.LoDTensor(lod_py)
+
+        lod_tensor.set_dims([5, 2, 3, 4])
+        lod_tensor.alloc_float(place)
+        tensor_array = numpy.array(lod_tensor)
         tensor_array[0, 0, 0, 0] = 1.0
         tensor_array[0, 0, 0, 1] = 2.0
-        tensor.set(tensor_array, place)
-        lod_py = [[0, 2, 5], [0, 2, 4, 5]]
+        lod_tensor.set(tensor_array, place)
 
-        lod_tensor = core.LoDTensor(lod_py, tensor)
-        lod_v = numpy.array(lod_tensor.tensor())
+        lod_v = numpy.array(lod_tensor)
         self.assertAlmostEqual(1.0, lod_v[0, 0, 0, 0])
         self.assertAlmostEqual(2.0, lod_v[0, 0, 0, 1])
         self.assertListEqual(lod_py, lod_tensor.lod())
diff --git a/python/paddle/v2/framework/tests/test_top_k_op.py b/python/paddle/v2/framework/tests/test_top_k_op.py
index cab799256d791889c295aa7f9048080f5caaf2dc..694f37d612d4c46e673dc894b05a0a446190732c 100644
--- a/python/paddle/v2/framework/tests/test_top_k_op.py
+++ b/python/paddle/v2/framework/tests/test_top_k_op.py
@@ -21,6 +21,9 @@ class TestTopkOp(OpTest):
 
         self.outputs = {'Out': output, 'Indices': indices}
 
+    def test_check_output(self):
+        self.check_output()
+
 
 class TestTopkOp3d(OpTest):
     def setUp(self):
@@ -42,6 +45,9 @@ class TestTopkOp3d(OpTest):
 
         self.outputs = {'Out': output, 'Indices': indices}
 
+    def test_check_output(self):
+        self.check_output()
+
 
 if __name__ == "__main__":
     unittest.main()
diff --git a/python/paddle/v2/framework/tests/test_uniform_random_op.py b/python/paddle/v2/framework/tests/test_uniform_random_op.py
index 76a5e36e56ab08230bdc2597d209fcf5d1d2acb0..9e8898fb5920defdfaa361bf45def7666a88beea 100644
--- a/python/paddle/v2/framework/tests/test_uniform_random_op.py
+++ b/python/paddle/v2/framework/tests/test_uniform_random_op.py
@@ -4,7 +4,7 @@ import paddle.v2.framework.core as core
 import numpy
 
 
-class UniformRandomTest(unittest.TestCase):
+class TestUniformRandomOp(unittest.TestCase):
     def test_uniform_random_cpu(self):
         self.uniform_random_test(place=core.CPUPlace())